{
  "id": 583023,
  "title": "What are your IMU-Only Scores",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/583023",
  "author_name": "Gerwyn",
  "post_date": "2025-06-04T09:47:24.209000",
  "votes": 45,
  "comment_count": 155,
  "views": 0,
  "content": "<p>Mine currently:</p>\n<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.72109</td>\n<td>0.73</td>\n</tr>\n<tr>\n<td>0.72974</td>\n<td>0.72</td>\n</tr>\n<tr>\n<td>0.74759</td>\n<td>0.74</td>\n</tr>\n<tr>\n<td>0.76612</td>\n<td>0.75</td>\n</tr>\n</tbody>\n</table>\n<ul>\n<li>CV using group on subject - stratified - 5 folds. (OOF vs CV-Average-Scores across 5-folds are very similar)</li>\n<li>LB is 5 folds predictions into majority voting.</li>\n</ul>\n<p>I am curious what are your experience with IMU-only performance</p>\n<p>edit 1: LB 0.74 <br>\nedit 2: LB 0.75</p>",
  "messages": [
    {
      "id": 3216928,
      "postDate": "2025-06-04T09:47:24.210Z",
      "content": "<p>Mine currently:</p>\n<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.72109</td>\n<td>0.73</td>\n</tr>\n<tr>\n<td>0.72974</td>\n<td>0.72</td>\n</tr>\n<tr>\n<td>0.74759</td>\n<td>0.74</td>\n</tr>\n<tr>\n<td>0.76612</td>\n<td>0.75</td>\n</tr>\n</tbody>\n</table>\n<ul>\n<li>CV using group on subject - stratified - 5 folds. (OOF vs CV-Average-Scores across 5-folds are very similar)</li>\n<li>LB is 5 folds predictions into majority voting.</li>\n</ul>\n<p>I am curious what are your experience with IMU-only performance</p>\n<p>edit 1: LB 0.74 <br>\nedit 2: LB 0.75</p>",
      "rawMarkdown": "Mine currently:\n\n|  cv | lb |\n| --- | --- |\n| 0.72109 | 0.73  |\n| 0.72974 | 0.72 |\n| 0.74759 | 0.74 |\n|0.76612 | 0.75 |\n\n- CV using group on subject - stratified - 5 folds. (OOF vs CV-Average-Scores across 5-folds are very similar)\n- LB is 5 folds predictions into majority voting.\n\n\nI am curious what are your experience with IMU-only performance\n\n\nedit 1: LB 0.74 \nedit 2: LB 0.75\n",
      "votes": 45
    },
    {
      "id": 3271800,
      "postDate": "2025-08-19T16:36:27.593Z",
      "content": "<p>CV: 0.845<br>\nLB: 0.831</p>",
      "rawMarkdown": "CV: 0.845\nLB: 0.831",
      "votes": 4,
      "replies": [
        {
          "id": 3271996,
          "postDate": "2025-08-20T04:22:18.563Z",
          "content": "<p>That's impressive! I can't get past 0.8 CV. Would you suggest focusing on a better model or deeper feature engineering?</p>",
          "rawMarkdown": "That's impressive! I can't get past 0.8 CV. Would you suggest focusing on a better model or deeper feature engineering?"
        },
        {
          "id": 3272027,
          "postDate": "2025-08-20T06:17:54.210Z",
          "content": "<p>Amazing CV！</p>",
          "rawMarkdown": "Amazing CV！"
        },
        {
          "id": 3272073,
          "postDate": "2025-08-20T08:08:12.913Z",
          "content": "<p>Amazing CV！My only IMU CV ~0.81 😂</p>",
          "rawMarkdown": "Amazing CV！My only IMU CV ~0.81 😂"
        },
        {
          "id": 3277621,
          "postDate": "2025-08-28T11:57:39.210Z",
          "content": "<p><a href=\"https://www.kaggle.com/l0glikelihood\" target=\"_blank\">@l0glikelihood</a> , is it just imu only ?  (CV: 0.845<br>\nLB: 0.831), no th and tof  ?</p>",
          "rawMarkdown": "@l0glikelihood , is it just imu only ?  (CV: 0.845\nLB: 0.831), no th and tof  ?",
          "replies": [
            {
              "id": 3277690,
              "postDate": "2025-08-28T14:03:46.360Z",
              "content": "<p>yes, just imu</p>",
              "rawMarkdown": "yes, just imu",
              "votes": 1
            }
          ]
        }
      ]
    },
    {
      "id": 3218857,
      "postDate": "2025-06-06T20:04:34.097Z",
      "content": "<p>update: <br>\nIMU only, 1d cnn<br>\ncv: 0.7782 lb:0.76</p>",
      "rawMarkdown": "update: \nIMU only, 1d cnn\ncv: 0.7782 lb:0.76",
      "votes": 4,
      "replies": [
        {
          "id": 3218880,
          "postDate": "2025-06-06T20:48:12.607Z",
          "content": "<p>If not sequence modelling, is this just with aggregate pre-processing / FE? Very impressive.</p>",
          "rawMarkdown": "If not sequence modelling, is this just with aggregate pre-processing / FE? Very impressive.",
          "votes": 1,
          "replies": [
            {
              "id": 3218888,
              "postDate": "2025-06-06T21:07:50.070Z",
              "content": "<p>I think 1D-CNN qualifies as a sequence model, as the convolutional kernel captures features along the time dimension, despite its limited time window. At this stage, I have applied some augmentation and feature engineering to the IMU data.</p>",
              "rawMarkdown": "I think 1D-CNN qualifies as a sequence model, as the convolutional kernel captures features along the time dimension, despite its limited time window. At this stage, I have applied some augmentation and feature engineering to the IMU data.",
              "votes": 4
            },
            {
              "id": 3218903,
              "postDate": "2025-06-06T22:15:58.853Z",
              "content": "<p>I've not thought about it that way before but I'd agree there, and I can see how a short time window could be effective for finding the pinches, scratches and pulls. I've been using 1D convolution with a dedicated LSTM subnet so far, but maybe I just need to find a more appropriate representation for the IMU data.</p>",
              "rawMarkdown": "I've not thought about it that way before but I'd agree there, and I can see how a short time window could be effective for finding the pinches, scratches and pulls. I've been using 1D convolution with a dedicated LSTM subnet so far, but maybe I just need to find a more appropriate representation for the IMU data."
            }
          ]
        },
        {
          "id": 3220421,
          "postDate": "2025-06-09T09:21:26.083Z",
          "content": "<p>update: cv: 0.7916 lb:0.78</p>",
          "rawMarkdown": "update: cv: 0.7916 lb:0.78",
          "votes": 3,
          "replies": [
            {
              "id": 3220440,
              "postDate": "2025-06-09T10:09:07.400Z",
              "content": "<p>That's nice, I get 0.787 or so but lb low 0.77. Do you do call back based on comp metric? I do and I am not sure if it inflates the 5-fold mean.</p>",
              "rawMarkdown": "That's nice, I get 0.787 or so but lb low 0.77. Do you do call back based on comp metric? I do and I am not sure if it inflates the 5-fold mean.",
              "votes": 1
            },
            {
              "id": 3220448,
              "postDate": "2025-06-09T10:22:47.933Z",
              "content": "<p>I am also monitoring the competition metric during training. In my experiment, CV-LB bias is around approximately -0.005 ~ -0.015. Therefore, I believe the CV inflation is not too severe.</p>",
              "rawMarkdown": "I am also monitoring the competition metric during training. In my experiment, CV-LB bias is around approximately -0.005 ~ -0.015. Therefore, I believe the CV inflation is not too severe.",
              "votes": 1
            },
            {
              "id": 3220450,
              "postDate": "2025-06-09T10:36:23.920Z",
              "content": "<p>Thanks for sharing that, it is in line with my observations as well! I meant that I restore weights of best val comp metric and I sometimes see little spikes here and there (imagine we restore a high outlier), which is why I believe correlation can get decoupled. Something like incorporating EMA should smooth it out I believe.</p>",
              "rawMarkdown": "Thanks for sharing that, it is in line with my observations as well! I meant that I restore weights of best val comp metric and I sometimes see little spikes here and there (imagine we restore a high outlier), which is why I believe correlation can get decoupled. Something like incorporating EMA should smooth it out I believe.",
              "votes": 1
            },
            {
              "id": 3221362,
              "postDate": "2025-06-10T20:55:41.307Z",
              "content": "<p>In my case I got cv 0.758 / lb 0.77 which is a consistent +0.01 on lb that was present in former submissions. So I might have the opposite problem? Using stratifiedgroupkfold like mostly everybody else I guess… 🤔</p>",
              "rawMarkdown": "In my case I got cv 0.758 / lb 0.77 which is a consistent +0.01 on lb that was present in former submissions. So I might have the opposite problem? Using stratifiedgroupkfold like mostly everybody else I guess... 🤔"
            },
            {
              "id": 3221378,
              "postDate": "2025-06-10T22:15:55.467Z",
              "content": "<p>I'm still using stratified fold instead of grouping by subject, so some degree of data leakage is expected in my results. By the way, would you mind sharing what target you use for StratifiedGroupKFold? I'm still uncertain about which label would be appropriate for stratified sampling, given that each subject has many sequences.</p>",
              "rawMarkdown": "I'm still using stratified fold instead of grouping by subject, so some degree of data leakage is expected in my results. By the way, would you mind sharing what target you use for StratifiedGroupKFold? I'm still uncertain about which label would be appropriate for stratified sampling, given that each subject has many sequences."
            },
            {
              "id": 3221554,
              "postDate": "2025-06-11T07:06:52.637Z",
              "content": "<p>So far I use this StratifiedGroupKFold using sequence_type, handedness, and sex as a combined target. Grouped by subject. What do you think about that?</p>\n<p>Best,<br>\nJan</p>",
              "rawMarkdown": "So far I use this StratifiedGroupKFold using sequence_type, handedness, and sex as a combined target. Grouped by subject. What do you think about that?\n\nBest,\nJan",
              "votes": 1
            },
            {
              "id": 3221559,
              "postDate": "2025-06-11T07:18:18.853Z",
              "content": "<p>That's an interesting approach, fiddling around with various label combinations is something I've tried as well. However, combining sequence_type ('Target' vs. 'Non-Target') is what you mean or that you use the gesture column? I played around with a concat of gesture - orientation and collapsing or not collapsing all non-target gestures into 1 class.</p>",
              "rawMarkdown": "That's an interesting approach, fiddling around with various label combinations is something I've tried as well. However, combining sequence_type ('Target' vs. 'Non-Target') is what you mean or that you use the gesture column? I played around with a concat of gesture - orientation and collapsing or not collapsing all non-target gestures into 1 class.",
              "votes": 1
            },
            {
              "id": 3221596,
              "postDate": "2025-06-11T08:15:22.900Z",
              "content": "<p>Thanks! I'll explore these targets and other possible combinations. Based on my intuition, handedness seems to be the most crucial feature for the split.</p>",
              "rawMarkdown": "Thanks! I'll explore these targets and other possible combinations. Based on my intuition, handedness seems to be the most crucial feature for the split.",
              "votes": 1
            },
            {
              "id": 3221844,
              "postDate": "2025-06-11T13:51:28.680Z",
              "content": "<p>Yes you are right. I did not yet fiddle with that stuff you mentioned but it would be interesting to hear about the results! 😊</p>",
              "rawMarkdown": "Yes you are right. I did not yet fiddle with that stuff you mentioned but it would be interesting to hear about the results! 😊"
            }
          ]
        },
        {
          "id": 3221650,
          "postDate": "2025-06-11T10:12:58.473Z",
          "content": "<p>update:<br>\nusing (gender + handedness + adult_child) as the target, StratifiedGroupKFold by subject has yielded a surprising result (experiment under same pipe):</p>\n<table>\n<thead>\n<tr>\n<th>Method</th>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>StratGroupKFold</td>\n<td>0.7619</td>\n<td>0.78</td>\n</tr>\n<tr>\n<td>Simple StratKFold using label</td>\n<td>0.7987</td>\n<td>0.78</td>\n</tr>\n</tbody>\n</table>\n<p>Interestingly, in my notebook version, StratGroupKFold's 0.78 appears higher than StratKFold's 0.78, despite both showing the same leaderboard score.</p>\n<p><a href=\"https://www.kaggle.com/janbrederecke\" target=\"_blank\">@janbrederecke</a> thx for the info!</p>",
          "rawMarkdown": "update:\nusing (gender + handedness + adult_child) as the target, StratifiedGroupKFold by subject has yielded a surprising result (experiment under same pipe):\n\n| Method | cv | lb |\n| --- | --- | --- |\n| StratGroupKFold | 0.7619 | 0.78 |\n| Simple StratKFold using label | 0.7987 | 0.78 |\n\nInterestingly, in my notebook version, StratGroupKFold's 0.78 appears higher than StratKFold's 0.78, despite both showing the same leaderboard score.\n\n@janbrederecke thx for the info!",
          "votes": 5,
          "replies": [
            {
              "id": 3221842,
              "postDate": "2025-06-11T13:50:04.193Z",
              "content": "<p>That is exactly what I found. Thanks for sharing your results as well! 🙂 </p>",
              "rawMarkdown": "That is exactly what I found. Thanks for sharing your results as well! 🙂 "
            },
            {
              "id": 3222109,
              "postDate": "2025-06-11T20:20:30.610Z",
              "content": "<p>Tried the experiment as well, 0.781 cv (sgkf, group on subject, target on handedness-sex-adult_child concat) and 0.79 lb, same lb as using stratified k fold on gesture (which had 0.8 cv).</p>",
              "rawMarkdown": "Tried the experiment as well, 0.781 cv (sgkf, group on subject, target on handedness-sex-adult_child concat) and 0.79 lb, same lb as using stratified k fold on gesture (which had 0.8 cv).",
              "votes": 1
            },
            {
              "id": 3222129,
              "postDate": "2025-06-11T21:30:37.167Z",
              "content": "<p>So does this prevent overfitting, I am not yet 100% sure…</p>",
              "rawMarkdown": "So does this prevent overfitting, I am not yet 100% sure..."
            },
            {
              "id": 3226589,
              "postDate": "2025-06-17T21:01:42.223Z",
              "content": "<p>What do you make of this? Would be interested in you perspective on these results.</p>",
              "rawMarkdown": "What do you make of this? Would be interested in you perspective on these results."
            },
            {
              "id": 3227137,
              "postDate": "2025-06-18T15:12:18.783Z",
              "content": "<p>Through further experiments, I found that the CV-LB correlation is not ideal for the group-by-subject method, especially in cases where CV increases but LB decreases. I believe this approach leads to more inconsistent results across folds, which may be reflected in the average CV score. In contrast, a simple stratified split yields a more stable and consistent correlation for me, even though facing some leakage in demo data.</p>",
              "rawMarkdown": "Through further experiments, I found that the CV-LB correlation is not ideal for the group-by-subject method, especially in cases where CV increases but LB decreases. I believe this approach leads to more inconsistent results across folds, which may be reflected in the average CV score. In contrast, a simple stratified split yields a more stable and consistent correlation for me, even though facing some leakage in demo data.",
              "votes": 3
            },
            {
              "id": 3227216,
              "postDate": "2025-06-18T17:00:51.383Z",
              "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> So you moved away from the gender + handedness + adult_child StratifiedGroupKFold in favor of StratKFold on label? I've noticed that improvements on cv with the gender + handedness + adult_child split haven't translated to lb. </p>",
              "rawMarkdown": "@ryenhails So you moved away from the gender + handedness + adult_child StratifiedGroupKFold in favor of StratKFold on label? I've noticed that improvements on cv with the gender + handedness + adult_child split haven't translated to lb. ",
              "votes": 1
            },
            {
              "id": 3227305,
              "postDate": "2025-06-18T19:47:39.477Z",
              "content": "<p>Yes, and I find it quite difficult to draw a conclusion. Even with the same CV split, the LB for a 0.84 CV is higher than that for 0.855.</p>",
              "rawMarkdown": "Yes, and I find it quite difficult to draw a conclusion. Even with the same CV split, the LB for a 0.84 CV is higher than that for 0.855.",
              "votes": 1
            },
            {
              "id": 3227306,
              "postDate": "2025-06-18T19:52:14.903Z",
              "content": "<p>I also find it hard, I get very similar LB results regardless of the split method, although regular stratified seems to have a lower std on 5 fold compared to gkfold. There's only 81 subjects so it's bound to be noisy. As to which one to trust over the other, don't really know, I get slightly better results with the regular stratified k fold on gesture.</p>",
              "rawMarkdown": "I also find it hard, I get very similar LB results regardless of the split method, although regular stratified seems to have a lower std on 5 fold compared to gkfold. There's only 81 subjects so it's bound to be noisy. As to which one to trust over the other, don't really know, I get slightly better results with the regular stratified k fold on gesture.",
              "votes": 2
            },
            {
              "id": 3227310,
              "postDate": "2025-06-18T20:01:16.017Z",
              "content": "<p>More or less the same here. And I hoped you guys could give me a definitive answer for this 😅. Thanks for sharing, appreciate it! 🙂</p>",
              "rawMarkdown": "More or less the same here. And I hoped you guys could give me a definitive answer for this 😅. Thanks for sharing, appreciate it! 🙂"
            },
            {
              "id": 3227318,
              "postDate": "2025-06-18T20:37:31.373Z",
              "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> <a href=\"https://www.kaggle.com/andreizamfir\" target=\"_blank\">@andreizamfir</a> <a href=\"https://www.kaggle.com/janbrederecke\" target=\"_blank\">@janbrederecke</a> Alright, since we're all active, who wants to share how preprocessing has impacted their results? Scaling by segment, by subject, by similar subjects, and global. </p>\n<p>My best submission was my first. No feature engineering, no scaling, just 1D-CNN action. A little perplexing seeing as research related to this problem always entails some scaling. </p>",
              "rawMarkdown": "@ryenhails @andreizamfir @janbrederecke Alright, since we're all active, who wants to share how preprocessing has impacted their results? Scaling by segment, by subject, by similar subjects, and global. \n\nMy best submission was my first. No feature engineering, no scaling, just 1D-CNN action. A little perplexing seeing as research related to this problem always entails some scaling. ",
              "votes": 1
            },
            {
              "id": 3227328,
              "postDate": "2025-06-18T20:48:33.010Z",
              "content": "<p>I am just using StandardScaler(), but I will try no scaling</p>",
              "rawMarkdown": "I am just using StandardScaler(), but I will try no scaling",
              "votes": 2
            },
            {
              "id": 3227360,
              "postDate": "2025-06-18T22:32:50.173Z",
              "content": "<p>I have to admit it is exactly the same for me - even though I tried a lot since then but no kind of preprocessing has helped so far. Would be really interested to find something that helps though!<br>\n<a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> , did you use standard scaler per sequence?</p>",
              "rawMarkdown": "I have to admit it is exactly the same for me - even though I tried a lot since then but no kind of preprocessing has helped so far. Would be really interested to find something that helps though!\n@ryenhails , did you use standard scaler per sequence?"
            },
            {
              "id": 3227363,
              "postDate": "2025-06-18T22:43:11.173Z",
              "content": "<p>I do absolutely no scaling at the moment, but it’s something I have on my list to revisit. I remember scaling in the beginning (per sequence), and seeing improvements after removing that and I just kept going as is since then.</p>\n<p>I do, however, employ FE and augmentations, both of which improved score in my case.</p>",
              "rawMarkdown": "I do absolutely no scaling at the moment, but it’s something I have on my list to revisit. I remember scaling in the beginning (per sequence), and seeing improvements after removing that and I just kept going as is since then.\n\nI do, however, employ FE and augmentations, both of which improved score in my case.",
              "votes": 1
            },
            {
              "id": 3227364,
              "postDate": "2025-06-18T22:47:48.147Z",
              "content": "<p><a href=\"https://www.kaggle.com/andreizamfir\" target=\"_blank\">@andreizamfir</a> would you mind sharing if you use 1d cnn here as well?</p>",
              "rawMarkdown": "@andreizamfir would you mind sharing if you use 1d cnn here as well?"
            },
            {
              "id": 3227365,
              "postDate": "2025-06-18T22:51:04.173Z",
              "content": "<p>Yeah, 1D CNN, arch changes provided biggest boosts in score. Think there’s a lot more to squeeze still</p>",
              "rawMarkdown": "Yeah, 1D CNN, arch changes provided biggest boosts in score. Think there’s a lot more to squeeze still",
              "votes": 1
            },
            {
              "id": 3227645,
              "postDate": "2025-06-19T06:43:20.633Z",
              "content": "<p><a href=\"https://www.kaggle.com/janbrederecke\" target=\"_blank\">@janbrederecke</a> yes, per sequence</p>",
              "rawMarkdown": "@janbrederecke yes, per sequence",
              "votes": 1
            },
            {
              "id": 3227676,
              "postDate": "2025-06-19T07:09:25.267Z",
              "content": "<p>Thanks guys, appreciate the open discussion in this comp! 😊</p>",
              "rawMarkdown": "Thanks guys, appreciate the open discussion in this comp! 😊",
              "votes": 1
            },
            {
              "id": 3235166,
              "postDate": "2025-06-28T19:25:08.740Z",
              "rawMarkdown": "",
              "isDeleted": true
            },
            {
              "id": 3235382,
              "postDate": "2025-06-29T06:41:19.467Z",
              "content": "<p>no gesture used. In my view, every subjects have all gestures, therefore it is not informative to group the subjects using gestures</p>",
              "rawMarkdown": "no gesture used. In my view, every subjects have all gestures, therefore it is not informative to group the subjects using gestures",
              "votes": 1
            }
          ]
        },
        {
          "id": 3227784,
          "postDate": "2025-06-19T09:17:38.260Z",
          "content": "<p>Update: <br>\nremoved StandardScaler() applied per sequence and the standardization of the demo column. This led to a improvement in LB score, but also introduced more cv-lb gap / data leakage when using a simple stratified split. <br>\nCV: 0.8022 → 0.8481 <br>\nLB: 0.79 → 0.80</p>",
          "rawMarkdown": "Update: \nremoved StandardScaler() applied per sequence and the standardization of the demo column. This led to a improvement in LB score, but also introduced more cv-lb gap / data leakage when using a simple stratified split. \nCV: 0.8022 → 0.8481 \nLB: 0.79 → 0.80",
          "votes": 2,
          "replies": [
            {
              "id": 3227786,
              "postDate": "2025-06-19T09:18:50.977Z",
              "content": "<p>Why do you call it data leakage? I am also wondering what that cv lb gap is exactly that is why I ask. 😊</p>",
              "rawMarkdown": "Why do you call it data leakage? I am also wondering what that cv lb gap is exactly that is why I ask. 😊"
            },
            {
              "id": 3227793,
              "postDate": "2025-06-19T09:30:16.083Z",
              "content": "<p>I believe the gap between the CV and LB scores can largely be attributed to two factors. First, external factors: differences in distributions between the test and train sets. Second, internal factors: the cross-validation split strategy within the train set, which can lead to data leakage from the training data into the validation folds.</p>\n<p>In my view, removing the StandardScaler() has minimal impact since it only scales information intrinsic to each sequence. In contrast, omitting the standardization of the demo column may lead to more severe data leakage. The scale of the demo column is significantly larger than that of the other features, which might cause the model to over-rely on these features for its predictions. Additionally, using a simple stratified split causes the same demo values to appear in both the training and validation sets, potentially widening the CV-LB gap.</p>\n<p>That said, this is merely my intuition—reality might prove otherwise.</p>",
              "rawMarkdown": "I believe the gap between the CV and LB scores can largely be attributed to two factors. First, external factors: differences in distributions between the test and train sets. Second, internal factors: the cross-validation split strategy within the train set, which can lead to data leakage from the training data into the validation folds.\n\nIn my view, removing the StandardScaler() has minimal impact since it only scales information intrinsic to each sequence. In contrast, omitting the standardization of the demo column may lead to more severe data leakage. The scale of the demo column is significantly larger than that of the other features, which might cause the model to over-rely on these features for its predictions. Additionally, using a simple stratified split causes the same demo values to appear in both the training and validation sets, potentially widening the CV-LB gap.\n\nThat said, this is merely my intuition—reality might prove otherwise."
            },
            {
              "id": 3227796,
              "postDate": "2025-06-19T09:40:45.453Z",
              "content": "<p>The data leakage when using demo data between training and validation is a very valid point. That’s where group k folding could avoid this sort of leakage.</p>\n<p>I think it’s part of the reason that you see this gap widening. </p>\n<p>I am not using demo data at all and my cv-lb is 0.84-0.82 on IMU data; this gap was consistent for me for the past 2 weeks.</p>",
              "rawMarkdown": "The data leakage when using demo data between training and validation is a very valid point. That’s where group k folding could avoid this sort of leakage.\n\nI think it’s part of the reason that you see this gap widening. \n\nI am not using demo data at all and my cv-lb is 0.84-0.82 on IMU data; this gap was consistent for me for the past 2 weeks.",
              "votes": 2
            },
            {
              "id": 3227804,
              "postDate": "2025-06-19T09:53:07.947Z",
              "content": "<p>Thanks, I will try no demo training</p>",
              "rawMarkdown": "Thanks, I will try no demo training"
            },
            {
              "id": 3227815,
              "postDate": "2025-06-19T10:07:49.853Z",
              "content": "<p>just for my understanding: no demo means, you just do group k fold using subject as the group and gesture as y? And for deeper understanding, the idea is that stratification introduces a kind of leakage into all the folds? Thanks so much for sharing I think this is a great perspective, will also try without if this is what you meant. </p>",
              "rawMarkdown": "just for my understanding: no demo means, you just do group k fold using subject as the group and gesture as y? And for deeper understanding, the idea is that stratification introduces a kind of leakage into all the folds? Thanks so much for sharing I think this is a great perspective, will also try without if this is what you meant. "
            },
            {
              "id": 3227830,
              "postDate": "2025-06-19T10:26:17.490Z",
              "content": "<p>So when I say I don’t use demo I meant I don’t use any sort of features based on the subject (age, sex, height etc) in training. I also regularly use StratifiedKFold based on gesture. </p>\n<p>Using StratifiedGroupKFold based on the subject as group and stratification based on some demographics still wouldn’t use demo features for training, it would just be a means of splitting data.</p>\n<p>The moment you do regular stratified k fold but also use demo features in training (feed them to whatever model you use), the situation described above by Ryenhails, you will have some of the same features in validation (as you’ll get a mixture of gestures from a particular subject in train and in val, and demo features are static and stricly subject dependent), hence the data leakage.</p>\n<p>(Above message is based on my understanding/intuition of what Ryenhails described, might be wrong!)</p>",
              "rawMarkdown": "So when I say I don’t use demo I meant I don’t use any sort of features based on the subject (age, sex, height etc) in training. I also regularly use StratifiedKFold based on gesture. \n\nUsing StratifiedGroupKFold based on the subject as group and stratification based on some demographics still wouldn’t use demo features for training, it would just be a means of splitting data.\n\nThe moment you do regular stratified k fold but also use demo features in training (feed them to whatever model you use), the situation described above by Ryenhails, you will have some of the same features in validation (as you’ll get a mixture of gestures from a particular subject in train and in val, and demo features are static and stricly subject dependent), hence the data leakage.\n\n(Above message is based on my understanding/intuition of what Ryenhails described, might be wrong!)",
              "votes": 1
            },
            {
              "id": 3227832,
              "postDate": "2025-06-19T10:28:20.017Z",
              "content": "<p>I believe that if the model does not use the demo columns, there would be no explicit leakage—since the same demo row won’t be split between training and validation, regardless of whether you use GroupKFold, stratified GroupKFold, or a simple stratified split. However, some latent leakage may still occur with a stratified split, as the same subjects could end up in both the training and validation sets.</p>",
              "rawMarkdown": "I believe that if the model does not use the demo columns, there would be no explicit leakage—since the same demo row won’t be split between training and validation, regardless of whether you use GroupKFold, stratified GroupKFold, or a simple stratified split. However, some latent leakage may still occur with a stratified split, as the same subjects could end up in both the training and validation sets."
            },
            {
              "id": 3227835,
              "postDate": "2025-06-19T10:35:09.710Z",
              "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> wouldn't this be prevented by using stratifiedgroupkfold and group by subject? </p>",
              "rawMarkdown": "@ryenhails wouldn't this be prevented by using stratifiedgroupkfold and group by subject? "
            },
            {
              "id": 3227837,
              "postDate": "2025-06-19T10:36:46.683Z",
              "content": "<p>That's exactly what I am thinking as well, that inherent leakage is there with regular stratified split, but it's more pronounced when you also use demographic features in training. Group K folding on subjects eliminates those particular leakage sources, although it causes noisy performance due to small number of subjects. Saving grace is the fact that each subject performed all gestures so you can maintain stratification of gestures with it.</p>",
              "rawMarkdown": "That's exactly what I am thinking as well, that inherent leakage is there with regular stratified split, but it's more pronounced when you also use demographic features in training. Group K folding on subjects eliminates those particular leakage sources, although it causes noisy performance due to small number of subjects. Saving grace is the fact that each subject performed all gestures so you can maintain stratification of gestures with it.",
              "votes": 1
            },
            {
              "id": 3227838,
              "postDate": "2025-06-19T10:43:42.273Z",
              "content": "<p>So what is our collective verdict rn? Use stratified group k fold with subject as the group and stratify by gesture? Or stratify by gesture + some other demographics? I have not used demographics in training so far and stratify by gesture and multiple demographics for now but experience the cv = lb + 0.01 discrepancy consistently.</p>",
              "rawMarkdown": "So what is our collective verdict rn? Use stratified group k fold with subject as the group and stratify by gesture? Or stratify by gesture + some other demographics? I have not used demographics in training so far and stratify by gesture and multiple demographics for now but experience the cv = lb + 0.01 discrepancy consistently."
            },
            {
              "id": 3227974,
              "postDate": "2025-06-19T14:05:51.500Z",
              "rawMarkdown": "",
              "votes": -1,
              "isDeleted": true
            },
            {
              "id": 3228012,
              "postDate": "2025-06-19T15:07:08.740Z",
              "content": "<p>for me, it is no demo, no group and use simple gesture stratify</p>",
              "rawMarkdown": "for me, it is no demo, no group and use simple gesture stratify",
              "votes": 1
            },
            {
              "id": 3228040,
              "postDate": "2025-06-19T15:38:53.327Z",
              "content": "<p>Appreciate it. Congrats on moving to 2nd! </p>",
              "rawMarkdown": "Appreciate it. Congrats on moving to 2nd! "
            },
            {
              "id": 3230146,
              "postDate": "2025-06-22T14:22:51.873Z",
              "content": "<p>impressive score! Was this achieved using just 7 IMU columns? Were orientation behavior data used as supplements?</p>",
              "rawMarkdown": "impressive score! Was this achieved using just 7 IMU columns? Were orientation behavior data used as supplements?"
            },
            {
              "id": 3230157,
              "postDate": "2025-06-22T14:35:26.337Z",
              "content": "<p>just 7 imu cols</p>",
              "rawMarkdown": "just 7 imu cols",
              "votes": 1
            },
            {
              "id": 3235535,
              "postDate": "2025-06-29T11:08:25.223Z",
              "content": "<p>Really? I cannot improve my LB score beyond 0.7</p>",
              "rawMarkdown": "Really? I cannot improve my LB score beyond 0.7"
            },
            {
              "id": 3239480,
              "postDate": "2025-07-02T21:15:03.800Z",
              "content": "<p>Wow, I am really excited to know, how many weights in your neural network? 😱</p>",
              "rawMarkdown": "Wow, I am really excited to know, how many weights in your neural network? 😱"
            },
            {
              "id": 3240189,
              "postDate": "2025-07-03T14:51:17.233Z",
              "content": "<p>model file ~100MB</p>",
              "rawMarkdown": "model file ~100MB",
              "votes": 2
            },
            {
              "id": 3241182,
              "postDate": "2025-07-04T16:05:12.260Z",
              "content": "<p>Is this per fold on IMU only?  </p>",
              "rawMarkdown": "Is this per fold on IMU only?  "
            },
            {
              "id": 3241332,
              "postDate": "2025-07-04T18:03:58.887Z",
              "content": "<p>yes per fold</p>",
              "rawMarkdown": "yes per fold"
            },
            {
              "id": 3241413,
              "postDate": "2025-07-04T19:16:14.520Z",
              "content": "<p>When you said iterate the arch, you meant it 😂That's over 3x my current fold sizes.</p>",
              "rawMarkdown": "When you said iterate the arch, you meant it 😂That's over 3x my current fold sizes."
            },
            {
              "id": 3241473,
              "postDate": "2025-07-04T20:23:07.613Z",
              "content": "<p>🤣🤣 I was just adding more and more blocks, but now it doesn’t seem to be improved anymore.</p>",
              "rawMarkdown": "🤣🤣 I was just adding more and more blocks, but now it doesn’t seem to be improved anymore."
            },
            {
              "id": 3241540,
              "postDate": "2025-07-04T21:30:50.263Z",
              "content": "<p><img src=\"https://tenor.com/view/star-wars-more-gif-21856591\" alt=\"MORE!\"></p>",
              "rawMarkdown": "![MORE!](https://tenor.com/view/star-wars-more-gif-21856591)"
            },
            {
              "id": 3241548,
              "postDate": "2025-07-04T21:38:58.483Z",
              "content": "<p>I’m increasing my model size and seeing CV improvements but I can’t see myself reaching 100MB lmao. Bro is out here 10 blocks deep.. I’ve got four blocks - any more and I overfit, even with lr warmup. </p>",
              "rawMarkdown": "I’m increasing my model size and seeing CV improvements but I can’t see myself reaching 100MB lmao. Bro is out here 10 blocks deep.. I’ve got four blocks - any more and I overfit, even with lr warmup. "
            },
            {
              "id": 3251244,
              "postDate": "2025-07-20T08:46:42.057Z",
              "rawMarkdown": "",
              "isDeleted": true
            }
          ]
        },
        {
          "id": 3229540,
          "postDate": "2025-06-21T16:27:13.957Z",
          "content": "<p>Very impressive, I'm wondering if you are also using the demo features for this, or just 1d cnn for the imu data?</p>",
          "rawMarkdown": "Very impressive, I'm wondering if you are also using the demo features for this, or just 1d cnn for the imu data?",
          "replies": [
            {
              "id": 3229561,
              "postDate": "2025-06-21T17:13:00.707Z",
              "content": "<p>no demo, only 1dcnn</p>",
              "rawMarkdown": "no demo, only 1dcnn"
            }
          ]
        },
        {
          "id": 3231367,
          "postDate": "2025-06-24T08:32:52.463Z",
          "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> hello, Ryenhails, Amazing lb. Is 0.78 still your highest score with only IMU now? If it's convenient.</p>",
          "rawMarkdown": "@ryenhails hello, Ryenhails, Amazing lb. Is 0.78 still your highest score with only IMU now? If it's convenient.",
          "replies": [
            {
              "id": 3231433,
              "postDate": "2025-06-24T10:34:19.127Z",
              "content": "<p>Hi, now my best IMU only cv is :0.8036(group by subject)/0.8522(without group), lb:0.81</p>",
              "rawMarkdown": "Hi, now my best IMU only cv is :0.8036(group by subject)/0.8522(without group), lb:0.81",
              "votes": 2
            },
            {
              "id": 3231455,
              "postDate": "2025-06-24T11:34:51.557Z",
              "content": "<p>Thank you, it confirms some of my ideas.</p>",
              "rawMarkdown": "Thank you, it confirms some of my ideas."
            },
            {
              "id": 3231642,
              "postDate": "2025-06-24T17:12:40.973Z",
              "content": "<p>I’m very curious how you’ve managed to separate eyebrow and eyelash hair pulls. I’ve got some ideas I’ll experiment with later… once I figure it out I’ll be right there with ya</p>",
              "rawMarkdown": "I’m very curious how you’ve managed to separate eyebrow and eyelash hair pulls. I’ve got some ideas I’ll experiment with later… once I figure it out I’ll be right there with ya"
            },
            {
              "id": 3231670,
              "postDate": "2025-06-24T18:03:04.553Z",
              "content": "<p>I have not check the specific class yet, just iterate the arch</p>",
              "rawMarkdown": "I have not check the specific class yet, just iterate the arch"
            }
          ]
        }
      ]
    },
    {
      "id": 3250014,
      "postDate": "2025-07-17T14:41:00.573Z",
      "content": "<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.8132</td>\n<td>0.82</td>\n</tr>\n<tr>\n<td>0.7968</td>\n<td>0.81</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "| CV | LB |\n| --- | --- |\n| 0.8132| 0.82 |\n| 0.7968 | 0.81 |\n",
      "votes": 1
    },
    {
      "id": 3225837,
      "postDate": "2025-06-16T22:56:27.597Z",
      "content": "<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.744</td>\n<td>0.75</td>\n</tr>\n<tr>\n<td>0.740</td>\n<td>0.76</td>\n</tr>\n<tr>\n<td>0.748</td>\n<td>0.75</td>\n</tr>\n<tr>\n<td>0.764</td>\n<td>0.77</td>\n</tr>\n</tbody>\n</table>\n<ul>\n<li>2nd 9 classes</li>\n<li>rest 18 classes</li>\n</ul>\n<p>CV: StratifiedGroupKFold -&gt; group=subject, target=gesture</p>",
      "rawMarkdown": "| CV | LB |\n| --- | --- |\n|0.744  |0.75  | \n|0.740  |0.76  |\n|0.748  |0.75  |\n|0.764  |0.77  |\n\n- 2nd 9 classes\n- rest 18 classes\n\nCV: StratifiedGroupKFold -> group=subject, target=gesture",
      "votes": 1
    },
    {
      "id": 3225814,
      "postDate": "2025-06-16T21:46:21.687Z",
      "content": "<p>cv - 0.7688<br>\nlb - 0.78 (mid)</p>",
      "rawMarkdown": "cv - 0.7688\nlb - 0.78 (mid)",
      "votes": 1,
      "replies": [
        {
          "id": 3249843,
          "postDate": "2025-07-17T08:06:32.273Z",
          "content": "<p>sgkf(group=subject, y=gesture, n_folds=5)</p>\n<p>cv - 0.826<br>\nlb - 0.82</p>",
          "rawMarkdown": "sgkf(group=subject, y=gesture, n_folds=5)\n\ncv - 0.826\nlb - 0.82",
          "votes": 1,
          "replies": [
            {
              "id": 3249852,
              "postDate": "2025-07-17T08:23:48.447Z",
              "content": "<p>Still IMU only? That's impressive</p>",
              "rawMarkdown": "Still IMU only? That's impressive"
            },
            {
              "id": 3249856,
              "postDate": "2025-07-17T08:34:15.993Z",
              "content": "<p>yes, imu only</p>",
              "rawMarkdown": "yes, imu only"
            },
            {
              "id": 3249982,
              "postDate": "2025-07-17T13:24:38.250Z",
              "content": "<p>You should be able to get top 30 right now with that. Features or architecture get you there? Or something else? </p>",
              "rawMarkdown": "You should be able to get top 30 right now with that. Features or architecture get you there? Or something else? "
            },
            {
              "id": 3249984,
              "postDate": "2025-07-17T13:33:16.450Z",
              "content": "<p>both of them</p>",
              "rawMarkdown": "both of them"
            },
            {
              "id": 3249995,
              "postDate": "2025-07-17T13:59:28.603Z",
              "content": "<p>I'm curious what features beyond public notebooks you've tried that were actually helpful. I've tried csums, diffs, longer ranged diffs, shifts, diffs from sequence median, wavelets, and some others. I even tried the 2d wavelets/spectograms as you did.</p>",
              "rawMarkdown": "I'm curious what features beyond public notebooks you've tried that were actually helpful. I've tried csums, diffs, longer ranged diffs, shifts, diffs from sequence median, wavelets, and some others. I even tried the 2d wavelets/spectograms as you did.",
              "votes": 2
            },
            {
              "id": 3249997,
              "postDate": "2025-07-17T14:02:41.837Z",
              "content": "<p>nothing other than what you listed</p>",
              "rawMarkdown": "nothing other than what you listed",
              "votes": 1
            },
            {
              "id": 3249998,
              "postDate": "2025-07-17T14:14:11.957Z",
              "content": "<p>And some of these were helpful for you? I've found my best set of features is even smaller than what most public notebooks have.. hmm</p>",
              "rawMarkdown": "And some of these were helpful for you? I've found my best set of features is even smaller than what most public notebooks have.. hmm"
            },
            {
              "id": 3261867,
              "postDate": "2025-08-02T11:56:23.777Z",
              "content": "<p>Did you do arch tuning? If you did then how much did it improve CV/LB?</p>",
              "rawMarkdown": "Did you do arch tuning? If you did then how much did it improve CV/LB?"
            }
          ]
        }
      ]
    },
    {
      "id": 3222454,
      "postDate": "2025-06-12T07:53:11.023Z",
      "content": "<p>using StratifiedGroupKFold (5fold, y is gesture, groups is subject)<br>\nOnly IMU:<br>\nCV: 0.795<br>\nLB: 0.75</p>\n<p>Amazing results…</p>",
      "rawMarkdown": "using StratifiedGroupKFold (5fold, y is gesture, groups is subject)\nOnly IMU:\nCV: 0.795\nLB: 0.75\n\nAmazing results...",
      "votes": 1
    },
    {
      "id": 3222022,
      "postDate": "2025-06-11T18:04:41.997Z",
      "content": "<p>cv-0.7583<br>\nlb-0.78<br>\nonly IMU</p>\n<p>update: cv-0.7714,lb-0.79</p>",
      "rawMarkdown": "cv-0.7583\nlb-0.78\nonly IMU\n\nupdate: cv-0.7714,lb-0.79",
      "votes": 1,
      "replies": [
        {
          "id": 3222534,
          "postDate": "2025-06-12T08:43:04.927Z",
          "content": "<p>What model did you use?</p>",
          "rawMarkdown": "What model did you use?\n",
          "replies": [
            {
              "id": 3222594,
              "postDate": "2025-06-12T09:34:58.760Z",
              "content": "<p>just 1d-cnn</p>",
              "rawMarkdown": "just 1d-cnn"
            },
            {
              "id": 3226584,
              "postDate": "2025-06-17T20:59:35.180Z",
              "content": "<p>Can you share how you create folds? Do you do stratification and if so only by the gestures or something else? I observe a similar cv / lb relationship using stratification.</p>",
              "rawMarkdown": "Can you share how you create folds? Do you do stratification and if so only by the gestures or something else? I observe a similar cv / lb relationship using stratification."
            }
          ]
        }
      ]
    },
    {
      "id": 3221623,
      "postDate": "2025-06-11T09:10:59.717Z",
      "content": "<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.762</td>\n<td>0.75</td>\n</tr>\n</tbody>\n</table>\n<p>5-fold catboost model<br>\nOut of ideas for now, time to look at the other sensors<br>\nAfter reading the comments here, might need to experiment with a different stratification for my cv too</p>",
      "rawMarkdown": "| CV | LB |\n| --- | --- |\n| 0.762 | 0.75 |\n\n5-fold catboost model\nOut of ideas for now, time to look at the other sensors\nAfter reading the comments here, might need to experiment with a different stratification for my cv too",
      "votes": 1
    },
    {
      "id": 3218831,
      "postDate": "2025-06-06T19:03:43.033Z",
      "content": "<p><a href=\"https://www.kaggle.com/gerwynng\" target=\"_blank\">@gerwynng</a> just starting out<br>\n<a href=\"https://www.kaggle.com/code/ravi20076/cmi2025-public-baseline-v3\" target=\"_blank\">Single catboost</a> CV 0.6904, LB = 0.69</p>",
      "rawMarkdown": "@gerwynng just starting out\n[Single catboost](https://www.kaggle.com/code/ravi20076/cmi2025-public-baseline-v3) CV 0.6904, LB = 0.69",
      "votes": 1
    },
    {
      "id": 3217902,
      "postDate": "2025-06-05T13:59:34.467Z",
      "content": "<p>CV = 0.723 (cv scores vary between 0.69-0.76 on 5 folds)<br>\nLB = 0.73 <br>\nUsing DL model with 1d CNN + LSTM. <br>\nProbably with an ensemble we can improve scores but now I will move to other sources of data.</p>\n<p>Edit:<br>\nLB = 0.75</p>\n<p>Edit 2:<br>\nSplit Stratified Group KFold (group=subject, target=sequence_type)<br>\nCV = 0.806, LB = 0.80 (mid)</p>",
      "rawMarkdown": "CV = 0.723 (cv scores vary between 0.69-0.76 on 5 folds)\nLB = 0.73 \nUsing DL model with 1d CNN + LSTM. \nProbably with an ensemble we can improve scores but now I will move to other sources of data.\n\nEdit:\nLB = 0.75\n\nEdit 2:\nSplit Stratified Group KFold (group=subject, target=sequence_type)\nCV = 0.806, LB = 0.80 (mid)",
      "votes": 1
    },
    {
      "id": 3217887,
      "postDate": "2025-06-05T13:39:07.683Z",
      "content": "<p>IMU Only, 2  1D CNN layers  1 dense layer starter:</p>\n<pre><code> of validation scores:\n   mean   : . ± .\n   binary : . ± .\n   macro  : . ± .\n</code></pre>\n<pre><code> : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n</code></pre>\n<p>L.E.: LB: 0.73</p>",
      "rawMarkdown": "IMU Only, 2  1D CNN layers  1 dense layer starter:\n\n```\nMean of validation scores:\n  F1 mean   : 0.7404 ± 0.0072\n  F1 binary : 0.9611 ± 0.0018\n  F1 macro  : 0.5196 ± 0.0135\n```\n\n```\nFold 0: F1 mean=0.7450, F1 binary=0.9642, F1 macro=0.5258\nFold 1: F1 mean=0.7409, F1 binary=0.9598, F1 macro=0.5220\nFold 2: F1 mean=0.7287, F1 binary=0.9590, F1 macro=0.4985\nFold 3: F1 mean=0.7371, F1 binary=0.9615, F1 macro=0.5127\nFold 4: F1 mean=0.7501, F1 binary=0.9611, F1 macro=0.5391\n```\n\nL.E.: LB: 0.73",
      "votes": 1,
      "replies": [
        {
          "id": 3217890,
          "postDate": "2025-06-05T13:42:04.347Z",
          "content": "<p>Impressive! what is your lb according to this cv</p>",
          "rawMarkdown": "Impressive! what is your lb according to this cv",
          "replies": [
            {
              "id": 3217896,
              "postDate": "2025-06-05T13:50:38.983Z",
              "content": "<p>Haven't sent it out yet, will probably do today and come back with an update.</p>",
              "rawMarkdown": "Haven't sent it out yet, will probably do today and come back with an update."
            }
          ]
        }
      ]
    },
    {
      "id": 3217700,
      "postDate": "2025-06-05T10:24:08.920Z",
      "content": "<p>only IMU 0.68</p>",
      "rawMarkdown": "only IMU 0.68",
      "votes": 1
    },
    {
      "id": 3217201,
      "postDate": "2025-06-04T16:48:09.417Z",
      "content": "<p>I did very few experiments so far and my current imu only scores are like this right now. I'll update this when I have some improvements.</p>\n<pre><code>  : ,\n  : ,\n  : \n</code></pre>",
      "rawMarkdown": "I did very few experiments so far and my current imu only scores are like this right now. I'll update this when I have some improvements.\n\n```python\n  \"f1_binary\": 0.9584695250913285,\n  \"f1_macro\": 0.4780790342101205,\n  \"f1_average\": 0.7182742796507245\n```",
      "votes": 1,
      "replies": [
        {
          "id": 3217271,
          "postDate": "2025-06-04T18:46:00.003Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 3217360,
          "postDate": "2025-06-04T23:58:13.763Z",
          "content": "<p>Im getting the same binary score, macro still around 0.36, average ~0.66. I dont know why Im having trouble breaking this threshold. Validation is stratified group kfold with 5 folds, using subject as group, no demographics whatsoever</p>",
          "rawMarkdown": "Im getting the same binary score, macro still around 0.36, average ~0.66. I dont know why Im having trouble breaking this threshold. Validation is stratified group kfold with 5 folds, using subject as group, no demographics whatsoever",
          "replies": [
            {
              "id": 3218051,
              "postDate": "2025-06-05T17:50:03.517Z",
              "rawMarkdown": "",
              "isDeleted": true
            }
          ]
        },
        {
          "id": 3217777,
          "postDate": "2025-06-05T11:50:51.087Z",
          "content": "<p>mine just hovers around my best cv:</p>\n<pre><code>\n     \n     \n     \n\n</code></pre>\n<p>usually <code>f1_binary</code> ~ 0.960, <code>f1_macro</code> ~0.49X</p>\n<p>would like to say that i have exhausted IMU so i can move on haha</p>",
          "rawMarkdown": "mine just hovers around my best cv:\n```json\n{\n    \"f1_binary\": 0.96254,\n    \"f1_macro\": 0.50305,\n    \"combined\": 0.73279\n}\n```\n\nusually `f1_binary` ~ 0.960, `f1_macro` ~0.49X\n\nwould like to say that i have exhausted IMU so i can move on haha",
          "votes": 2
        }
      ]
    },
    {
      "id": 3217166,
      "postDate": "2025-06-04T15:39:59.367Z",
      "content": "<table>\n<thead>\n<tr>\n<th>5folds cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.6980</td>\n<td>0.69</td>\n</tr>\n<tr>\n<td>0.7137</td>\n<td>0.71</td>\n</tr>\n<tr>\n<td>0.7329</td>\n<td>0.73</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "|5folds cv | lb |\n| --- | --- |\n| 0.6980 | 0.69 |\n| 0.7137 | 0.71 |\n| 0.7329 | 0.73 |\n",
      "votes": 1,
      "replies": [
        {
          "id": 3217779,
          "postDate": "2025-06-05T11:52:13.933Z",
          "content": "<p>ahh nice, seems to corroborate with my experience, that scores <em>start to</em> plateau around 0.73 with IMU only </p>",
          "rawMarkdown": "ahh nice, seems to corroborate with my experience, that scores *start to* plateau around 0.73 with IMU only "
        }
      ]
    },
    {
      "id": 3224688,
      "postDate": "2025-06-15T10:27:30.480Z",
      "content": "<p>Only IMU</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.728</td>\n<td>0.72</td>\n</tr>\n<tr>\n<td>0.748</td>\n<td>0.74</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "Only IMU\n\n| CV    | LB   |\n|-------|------|\n| 0.728 | 0.72 |\n| 0.748 | 0.74 |",
      "votes": 2,
      "replies": [
        {
          "id": 3229538,
          "postDate": "2025-06-21T16:25:14.890Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 3218596,
      "postDate": "2025-06-06T12:06:31.020Z",
      "content": "<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.724</td>\n<td>0.72</td>\n</tr>\n<tr>\n<td>0.751</td>\n<td>0.74</td>\n</tr>\n</tbody>\n</table>\n<p>only imu and default params for gbdt. still a lot to do.</p>",
      "rawMarkdown": "| cv | lb |\n| --- | --- |\n| 0.724 | 0.72 |\n| 0.751 | 0.74 |\n\n\nonly imu and default params for gbdt. still a lot to do.",
      "votes": 2,
      "replies": [
        {
          "id": 3221340,
          "postDate": "2025-06-10T19:40:32.830Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 3237027,
          "postDate": "2025-06-30T20:30:39.867Z",
          "content": "<p>Update:</p>\n<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.758</td>\n<td>0.75</td>\n</tr>\n<tr>\n<td>0.763</td>\n<td>0.75</td>\n</tr>\n</tbody>\n</table>\n<p>still only imu and GBDT, it's hard to beat CNN-based models, now I can move on to them.</p>\n<p>Update:</p>\n<p>switched to CNN, initial experiments:</p>\n<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.742</td>\n<td>0.77</td>\n</tr>\n</tbody>\n</table>",
          "rawMarkdown": "Update:\n\n| cv | lb |\n| --- | --- |\n| 0.758 | 0.75 |\n| 0.763 | 0.75 |\n\nstill only imu and GBDT, it's hard to beat CNN-based models, now I can move on to them.\n\nUpdate:\n\nswitched to CNN, initial experiments:\n\n| cv | lb |\n| --- | --- |\n| 0.742| 0.77|"
        }
      ]
    },
    {
      "id": 3218227,
      "postDate": "2025-06-06T02:10:20.537Z",
      "content": "<p>first trial 5-fold cv: 0.717 lb:0.7</p>",
      "rawMarkdown": "first trial 5-fold cv: 0.717 lb:0.7",
      "votes": 2,
      "replies": [
        {
          "id": 3220476,
          "postDate": "2025-06-09T11:41:01.820Z",
          "content": "<p>Update today:</p>\n<table>\n<thead>\n<tr>\n<th>Cv</th>\n<th>Lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.7085</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7099</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7109</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7121</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7195</td>\n<td>0.71</td>\n</tr>\n<tr>\n<td>0.7216</td>\n<td>0.72</td>\n</tr>\n</tbody>\n</table>",
          "rawMarkdown": "Update today:\n\n| Cv     | Lb   |\n| ------ | ---- |\n| 0.7085 | 0.7  |\n| 0.7099 | 0.7  |\n| 0.7109 | 0.7  |\n| 0.7121 | 0.7  |\n| 0.7195 | 0.71 |\n| 0.7216 | 0.72 |"
        }
      ]
    },
    {
      "id": 3217632,
      "postDate": "2025-06-05T08:22:05.343Z",
      "content": "<h4>My First Test with IMU Only</h4>\n<p>Here are the initial results:</p>\n<table>\n<thead>\n<tr>\n<th>5 Fold OOF CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.7052</td>\n<td>0.70</td>\n</tr>\n</tbody>\n</table>\n<p>Seem like a lot to improve upon…</p>",
      "rawMarkdown": "#### My First Test with IMU Only\n\nHere are the initial results:\n\n| 5 Fold OOF CV     | LB   |\n|--------|------|\n| 0.7052 | 0.70 |\n\nSeem like a lot to improve upon...",
      "votes": 2,
      "replies": [
        {
          "id": 3218587,
          "postDate": "2025-06-06T11:50:11.017Z",
          "content": "<table>\n<thead>\n<tr>\n<th>5 Fold OOF CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.7458</td>\n<td>0.74</td>\n</tr>\n</tbody>\n</table>\n<p>Still IMU only and no cnn or other seq models. Wonder what's the extend 🤔</p>",
          "rawMarkdown": "| 5 Fold OOF CV     | LB   |\n|--------|------|\n| 0.7458 | 0.74 |\n\nStill IMU only and no cnn or other seq models. Wonder what's the extend 🤔",
          "votes": 2,
          "replies": [
            {
              "id": 3218592,
              "postDate": "2025-06-06T11:53:50.080Z",
              "content": "<p>Are these results for GBM models?</p>",
              "rawMarkdown": "Are these results for GBM models?",
              "votes": 1
            },
            {
              "id": 3218594,
              "postDate": "2025-06-06T11:57:42.893Z",
              "content": "<p>i havent been able to surpass 0.73X, good to know 0.74 can be done 😃</p>",
              "rawMarkdown": "i havent been able to surpass 0.73X, good to know 0.74 can be done 😃",
              "votes": 1
            },
            {
              "id": 3218595,
              "postDate": "2025-06-06T11:58:24.830Z",
              "content": "<p><a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a> </p>\n<p>Yeah, almost with default hparams, not optimized at all</p>",
              "rawMarkdown": "@alejopaullier \n\nYeah, almost with default hparams, not optimized at all",
              "votes": 1
            }
          ]
        }
      ]
    },
    {
      "id": 3248648,
      "postDate": "2025-07-15T00:44:33.510Z",
      "content": "<p>cv - 0.7983<br>\nlb - 0.80 (10-fold)</p>",
      "rawMarkdown": "cv - 0.7983\nlb - 0.80 (10-fold)"
    },
    {
      "id": 3243986,
      "postDate": "2025-07-07T17:30:17.037Z",
      "content": "<p>StratifiedGroupKFold group=subject, y=gesture</p>\n<p>cv: 0.797<br>\nlb: 0.8</p>",
      "rawMarkdown": "StratifiedGroupKFold group=subject, y=gesture\n\ncv: 0.797\nlb: 0.8"
    },
    {
      "id": 3243928,
      "postDate": "2025-07-07T16:49:35.173Z",
      "content": "<p>1D CNN IMU only    (5 folds group stratified)</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.791</td>\n<td>0.78</td>\n</tr>\n<tr>\n<td>0.780</td>\n<td>0.77</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "1D CNN IMU only    (5 folds group stratified)\n| CV | LB |\n| --- | --- |\n| 0.791 |0.78 |\n| 0.780|0.77 |\n"
    },
    {
      "id": 3243876,
      "postDate": "2025-07-07T15:56:53.857Z",
      "content": "<p>Hi,<br>\nI did merge demographics and training data. Initially when I used all sensor data, the model performed very poorly (0.18) but when the IMU data alone was used (still merged with demographics) the accuracy improved (0.52). Still haven't used the test.csv though.</p>\n<p>But how to improve further..</p>\n<p>I have attached my notebook below..</p>",
      "rawMarkdown": "Hi,\nI did merge demographics and training data. Initially when I used all sensor data, the model performed very poorly (0.18) but when the IMU data alone was used (still merged with demographics) the accuracy improved (0.52). Still haven't used the test.csv though.\n\nBut how to improve further..\n\nI have attached my notebook below.."
    },
    {
      "id": 3242945,
      "postDate": "2025-07-06T15:10:42.410Z",
      "content": "<p>StratifiedGroupKFold, split by subject: cv : 0.7789 LB : 0.79</p>",
      "rawMarkdown": "StratifiedGroupKFold, split by subject: cv : 0.7789 LB : 0.79",
      "replies": [
        {
          "id": 3243358,
          "postDate": "2025-07-07T04:09:15.933Z",
          "content": "<p>test1 : CV : 0.7904 LB : 0.81 (low)</p>",
          "rawMarkdown": "test1 : CV : 0.7904 LB : 0.81 (low)",
          "replies": [
            {
              "id": 3245809,
              "postDate": "2025-07-09T22:19:24.323Z",
              "content": "<p>is that imu only？</p>",
              "rawMarkdown": "is that imu only？"
            }
          ]
        }
      ]
    },
    {
      "id": 3241015,
      "postDate": "2025-07-04T12:48:02.290Z",
      "content": "<p>I'm really curious whether existing a team approaches the information of tof and thm only using imu.</p>",
      "rawMarkdown": "I'm really curious whether existing a team approaches the information of tof and thm only using imu."
    },
    {
      "id": 3240995,
      "postDate": "2025-07-04T12:29:50.203Z",
      "content": "<p>update: just did my first NN model submission:</p>\n<p>cv - 0.7952, LB - 0.79</p>",
      "rawMarkdown": "update: just did my first NN model submission:\n\ncv - 0.7952, LB - 0.79"
    },
    {
      "id": 3240924,
      "postDate": "2025-07-04T10:49:30.690Z",
      "content": "<p>I am using StratifiedKFold with the variables sequence_id and gesture.<br>\ncv 0.8755 / LB 0.80 🤣</p>",
      "rawMarkdown": "I am using StratifiedKFold with the variables sequence_id and gesture.\ncv 0.8755 / LB 0.80 🤣",
      "replies": [
        {
          "id": 3240993,
          "postDate": "2025-07-04T12:27:48.843Z",
          "content": "<p>Wow, what a CV 😅. Anything to share about that? Kudos if that is correct! </p>",
          "rawMarkdown": "Wow, what a CV 😅. Anything to share about that? Kudos if that is correct! "
        }
      ]
    },
    {
      "id": 3240759,
      "postDate": "2025-07-04T08:01:13.083Z",
      "content": "<p>Only IMU - CV = 0.83; LB=0.81</p>",
      "rawMarkdown": "Only IMU - CV = 0.83; LB=0.81",
      "replies": [
        {
          "id": 3241422,
          "postDate": "2025-07-04T19:27:20.247Z",
          "content": "<p>What kind of model are you missing? If possible. I am stuck at 0.76 with 1d cnn imu only model</p>",
          "rawMarkdown": "What kind of model are you missing? If possible. I am stuck at 0.76 with 1d cnn imu only model"
        }
      ]
    },
    {
      "id": 3239538,
      "postDate": "2025-07-02T23:25:53.377Z",
      "content": "<p>CV:0.82, LB:0.78 </p>",
      "rawMarkdown": "CV:0.82, LB:0.78 ",
      "replies": [
        {
          "id": 3246563,
          "postDate": "2025-07-11T06:14:46.243Z",
          "content": "<p>update cv: 0.79, LB 0.79</p>",
          "rawMarkdown": "update cv: 0.79, LB 0.79"
        }
      ]
    },
    {
      "id": 3233103,
      "postDate": "2025-06-26T13:35:10.560Z",
      "content": "<table>\n<thead>\n<tr>\n<th><strong>Date</strong></th>\n<th><strong>5 Fold CV</strong></th>\n<th><strong>Public LB</strong></th>\n<th><strong>CV Strategy</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>26. June</td>\n<td>0.759124</td>\n<td>0.74</td>\n<td>StratifiedGroupKFold, groups=subject</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "| **Date** | **5 Fold CV** | **Public LB** |            **CV Strategy**           |\n|:--------:|:-------------:|:-------------:|:------------------------------------:|\n| 26. June |      0.759124     |      0.74     | StratifiedGroupKFold, groups=subject |"
    },
    {
      "id": 3232810,
      "postDate": "2025-06-26T08:37:09.647Z",
      "content": "<p>IMU + THM + TOF, StratifiedGroupKFold:</p>\n<ul>\n<li>CV: 0.8027. Public leader board: 0.81.</li>\n<li>CV: 0.8840. Public leader board: 0.81.<br>\n🤔😅</li>\n</ul>",
      "rawMarkdown": "IMU + THM + TOF, StratifiedGroupKFold:\n- CV: 0.8027. Public leader board: 0.81.\n- CV: 0.8840. Public leader board: 0.81.\n🤔😅",
      "replies": [
        {
          "id": 3233030,
          "postDate": "2025-06-26T12:33:06.850Z",
          "content": "<p>So it is not IMU only? 🤔</p>",
          "rawMarkdown": "So it is not IMU only? 🤔",
          "replies": [
            {
              "id": 3233949,
              "postDate": "2025-06-27T11:15:03.553Z",
              "content": "<p>I use all IMU + THM + TOF.</p>",
              "rawMarkdown": "I use all IMU + THM + TOF."
            }
          ]
        },
        {
          "id": 3244292,
          "postDate": "2025-07-08T04:14:07.023Z",
          "content": "<p>How did you manage your RAM limit?</p>",
          "rawMarkdown": "How did you manage your RAM limit?"
        }
      ]
    },
    {
      "id": 3230797,
      "postDate": "2025-06-23T12:56:51.077Z",
      "content": "<p>IMU only 5 folds group stratified (group=subject, y=sequence_type), no demographics.</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB (single fold)</th>\n<th>LB (5 fold avg)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.756</td>\n<td>0.74 (mid)</td>\n<td>0.75 (mid-high)</td>\n</tr>\n<tr>\n<td>0.783</td>\n<td>NaN</td>\n<td>0.78 (low)</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "IMU only 5 folds group stratified (group=subject, y=sequence_type), no demographics.\n\n| CV | LB (single fold) | LB (5 fold avg)|\n| --- | --- |---|\n| 0.756 | 0.74 (mid) | 0.75 (mid-high)|\n| 0.783 | NaN | 0.78 (low)|",
      "replies": [
        {
          "id": 3230827,
          "postDate": "2025-06-23T13:45:48.757Z",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/optimo\" target=\"_blank\">@optimo</a> IMU data means you are not considering \"tof_[1-5]_v[0-63]\" columns only columns related to acceleration and rotations ?</p>",
          "rawMarkdown": "Hi @optimo IMU data means you are not considering \"tof_[1-5]_v[0-63]\" columns only columns related to acceleration and rotations ?",
          "isDeleted": true,
          "replies": [
            {
              "id": 3230849,
              "postDate": "2025-06-23T14:30:18.377Z",
              "content": "<p>It means only features from acceleration and rotation can enter my model at the moment, no \"tof_\" and no \"thm_\".</p>",
              "rawMarkdown": "It means only features from acceleration and rotation can enter my model at the moment, no \"tof_\" and no \"thm_\"."
            }
          ]
        },
        {
          "id": 3246558,
          "postDate": "2025-07-11T06:09:17.067Z",
          "content": "<p>Can you plz update your imu only score if it is not secret?</p>",
          "rawMarkdown": "Can you plz update your imu only score if it is not secret?",
          "votes": -6
        }
      ]
    },
    {
      "id": 3230252,
      "postDate": "2025-06-22T16:10:18.350Z",
      "content": "<p>CV 0.728 LB 0.74</p>",
      "rawMarkdown": "CV 0.728 LB 0.74"
    },
    {
      "id": 3228985,
      "postDate": "2025-06-20T19:30:40.470Z",
      "content": "<p>cv: 0.7915, lb: 0.78</p>\n<p>update: <br>\ncv: 0.8121 lb: 0.8</p>\n<p>update:<br>\ncv: 0.8209, lb: 0.81</p>\n<p>lb: 0.82 </p>",
      "rawMarkdown": "cv: 0.7915, lb: 0.78\n\nupdate: \ncv: 0.8121 lb: 0.8\n\nupdate:\ncv: 0.8209, lb: 0.81\n\nlb: 0.82 ",
      "replies": [
        {
          "id": 3246565,
          "postDate": "2025-07-11T06:15:57.313Z",
          "content": "<p>May I ask how large is your model if that is not secret?</p>",
          "rawMarkdown": "May I ask how large is your model if that is not secret?",
          "replies": [
            {
              "id": 3246736,
              "postDate": "2025-07-11T12:34:39.763Z",
              "content": "<p>I have two architectures that have hit .81 with IMU only. One has a fold size of about 55MB, the other is 5.5MB. </p>",
              "rawMarkdown": "I have two architectures that have hit .81 with IMU only. One has a fold size of about 55MB, the other is 5.5MB. "
            },
            {
              "id": 3246758,
              "postDate": "2025-07-11T13:40:22.657Z",
              "content": "<p>Wow that's pretty small, I believe you did a good job in feature engineering 😃</p>",
              "rawMarkdown": "Wow that's pretty small, I believe you did a good job in feature engineering 😃"
            }
          ]
        }
      ]
    },
    {
      "id": 3227729,
      "postDate": "2025-06-19T08:21:25.057Z",
      "content": "<p>IMU ONLY:<br>\nCV:0.7738  LB:0.76<br>\nIMU+TOF:<br>\nCV:0.85  LB:0.79</p>",
      "rawMarkdown": "IMU ONLY:\nCV:0.7738  LB:0.76\nIMU+TOF:\nCV:0.85  LB:0.79",
      "replies": [
        {
          "id": 3227780,
          "postDate": "2025-06-19T09:09:26.310Z",
          "content": "<p>I use seq_id to split fold, my cv may be not right</p>",
          "rawMarkdown": "I use seq_id to split fold, my cv may be not right"
        }
      ]
    },
    {
      "id": 3225581,
      "postDate": "2025-06-16T15:46:14.017Z",
      "content": "<p>GroupKfold stratified based on subject and gesture<br>\nCV = 0.6742<br>\nLB = 0.68</p>",
      "rawMarkdown": "GroupKfold stratified based on subject and gesture\nCV = 0.6742\nLB = 0.68"
    },
    {
      "id": 3225197,
      "postDate": "2025-06-16T05:14:34.713Z",
      "content": "<p>Update:<br>\ncv - 0.76612<br>\nlb - 0.75</p>\n<p>I have several improvements (different enough experiments) that improved CV but did not translate into improvements on LB. I would think it's quite easy to overfit to a few sequence_ids, once performance is good enough. </p>",
      "rawMarkdown": "Update:\ncv - 0.76612\nlb - 0.75\n\n\nI have several improvements (different enough experiments) that improved CV but did not translate into improvements on LB. I would think it's quite easy to overfit to a few sequence_ids, once performance is good enough. "
    },
    {
      "id": 3218385,
      "postDate": "2025-06-06T05:57:26.213Z",
      "content": "<p><code>accuracy: 0.9753\n f1_macro: 0.9762 \n loss: 0.4299</code><br>\nLB: 0.69</p>",
      "rawMarkdown": "`accuracy: 0.9753\n f1_macro: 0.9762 \n loss: 0.4299`\nLB: 0.69"
    },
    {
      "id": 3217520,
      "postDate": "2025-06-05T05:16:05.563Z",
      "content": "<p>After doing groupby are you removing duplicates? aren't that reducing the data shape significantly?</p>",
      "rawMarkdown": "After doing groupby are you removing duplicates? aren't that reducing the data shape significantly?"
    },
    {
      "id": 3216977,
      "postDate": "2025-06-04T11:20:05.540Z",
      "content": "<p>May I ask if your score is using the raw sequences or did you apply any preprocessing besides padding, scaling, trimming?</p>",
      "rawMarkdown": "May I ask if your score is using the raw sequences or did you apply any preprocessing besides padding, scaling, trimming?",
      "replies": [
        {
          "id": 3216992,
          "postDate": "2025-06-04T11:41:09.630Z",
          "content": "<p>I compute aggregate statistic by sequence_id and fit with gbm</p>",
          "rawMarkdown": "I compute aggregate statistic by sequence_id and fit with gbm",
          "votes": 2
        }
      ]
    },
    {
      "id": 3228874,
      "postDate": "2025-06-20T16:03:22.177Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 3228550,
      "postDate": "2025-06-20T09:23:18.243Z",
      "content": "<p><a href=\"https://www.kaggle.com/gerwynng\" target=\"_blank\">@gerwynng</a> IMU data means you are not considering <strong>\"tof_[1-5]_v[0-63]\"</strong> columns ?</p>",
      "rawMarkdown": "@gerwynng IMU data means you are not considering **\"tof_[1-5]_v[0-63]\"** columns ?",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3271800,
      "author_name": "kss",
      "author_url": "",
      "post_date": "2025-08-19T16:36:27.593000",
      "content": "<p>CV: 0.845<br>\nLB: 0.831</p>",
      "votes": 4,
      "replies": [
        {
          "id": 3271996,
          "author_name": "EliKal",
          "author_url": "",
          "post_date": "2025-08-20T04:22:18.563000",
          "content": "<p>That's impressive! I can't get past 0.8 CV. Would you suggest focusing on a better model or deeper feature engineering?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3272027,
          "author_name": "Timmy Juicehouse",
          "author_url": "",
          "post_date": "2025-08-20T06:17:54.210000",
          "content": "<p>Amazing CV！</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3272073,
          "author_name": "Quan Vu",
          "author_url": "",
          "post_date": "2025-08-20T08:08:12.913000",
          "content": "<p>Amazing CV！My only IMU CV ~0.81 😂</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3277621,
          "author_name": "Manoj",
          "author_url": "",
          "post_date": "2025-08-28T11:57:39.210000",
          "content": "<p><a href=\"https://www.kaggle.com/l0glikelihood\" target=\"_blank\">@l0glikelihood</a> , is it just imu only ?  (CV: 0.845<br>\nLB: 0.831), no th and tof  ?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3277690,
              "author_name": "kss",
              "author_url": "",
              "post_date": "2025-08-28T14:03:46.360000",
              "content": "<p>yes, just imu</p>",
              "votes": 1,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3218857,
      "author_name": "Ryenhails",
      "author_url": "",
      "post_date": "2025-06-06T20:04:34.097000",
      "content": "<p>update: <br>\nIMU only, 1d cnn<br>\ncv: 0.7782 lb:0.76</p>",
      "votes": 4,
      "replies": [
        {
          "id": 3218880,
          "author_name": "Peter Moorhouse",
          "author_url": "",
          "post_date": "2025-06-06T20:48:12.607000",
          "content": "<p>If not sequence modelling, is this just with aggregate pre-processing / FE? Very impressive.</p>",
          "votes": 1,
          "replies": [
            {
              "id": 3218888,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-06T21:07:50.070000",
              "content": "<p>I think 1D-CNN qualifies as a sequence model, as the convolutional kernel captures features along the time dimension, despite its limited time window. At this stage, I have applied some augmentation and feature engineering to the IMU data.</p>",
              "votes": 4,
              "replies": []
            },
            {
              "id": 3218903,
              "author_name": "Peter Moorhouse",
              "author_url": "",
              "post_date": "2025-06-06T22:15:58.853000",
              "content": "<p>I've not thought about it that way before but I'd agree there, and I can see how a short time window could be effective for finding the pinches, scratches and pulls. I've been using 1D convolution with a dedicated LSTM subnet so far, but maybe I just need to find a more appropriate representation for the IMU data.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3220421,
          "author_name": "Ryenhails",
          "author_url": "",
          "post_date": "2025-06-09T09:21:26.083000",
          "content": "<p>update: cv: 0.7916 lb:0.78</p>",
          "votes": 3,
          "replies": [
            {
              "id": 3220440,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-09T10:09:07.400000",
              "content": "<p>That's nice, I get 0.787 or so but lb low 0.77. Do you do call back based on comp metric? I do and I am not sure if it inflates the 5-fold mean.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3220448,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-09T10:22:47.933000",
              "content": "<p>I am also monitoring the competition metric during training. In my experiment, CV-LB bias is around approximately -0.005 ~ -0.015. Therefore, I believe the CV inflation is not too severe.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3220450,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-09T10:36:23.920000",
              "content": "<p>Thanks for sharing that, it is in line with my observations as well! I meant that I restore weights of best val comp metric and I sometimes see little spikes here and there (imagine we restore a high outlier), which is why I believe correlation can get decoupled. Something like incorporating EMA should smooth it out I believe.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3221362,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-10T20:55:41.307000",
              "content": "<p>In my case I got cv 0.758 / lb 0.77 which is a consistent +0.01 on lb that was present in former submissions. So I might have the opposite problem? Using stratifiedgroupkfold like mostly everybody else I guess… 🤔</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3221378,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-10T22:15:55.467000",
              "content": "<p>I'm still using stratified fold instead of grouping by subject, so some degree of data leakage is expected in my results. By the way, would you mind sharing what target you use for StratifiedGroupKFold? I'm still uncertain about which label would be appropriate for stratified sampling, given that each subject has many sequences.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3221554,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-11T07:06:52.637000",
              "content": "<p>So far I use this StratifiedGroupKFold using sequence_type, handedness, and sex as a combined target. Grouped by subject. What do you think about that?</p>\n<p>Best,<br>\nJan</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3221559,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-11T07:18:18.853000",
              "content": "<p>That's an interesting approach, fiddling around with various label combinations is something I've tried as well. However, combining sequence_type ('Target' vs. 'Non-Target') is what you mean or that you use the gesture column? I played around with a concat of gesture - orientation and collapsing or not collapsing all non-target gestures into 1 class.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3221596,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-11T08:15:22.900000",
              "content": "<p>Thanks! I'll explore these targets and other possible combinations. Based on my intuition, handedness seems to be the most crucial feature for the split.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3221844,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-11T13:51:28.680000",
              "content": "<p>Yes you are right. I did not yet fiddle with that stuff you mentioned but it would be interesting to hear about the results! 😊</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3221650,
          "author_name": "Ryenhails",
          "author_url": "",
          "post_date": "2025-06-11T10:12:58.473000",
          "content": "<p>update:<br>\nusing (gender + handedness + adult_child) as the target, StratifiedGroupKFold by subject has yielded a surprising result (experiment under same pipe):</p>\n<table>\n<thead>\n<tr>\n<th>Method</th>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>StratGroupKFold</td>\n<td>0.7619</td>\n<td>0.78</td>\n</tr>\n<tr>\n<td>Simple StratKFold using label</td>\n<td>0.7987</td>\n<td>0.78</td>\n</tr>\n</tbody>\n</table>\n<p>Interestingly, in my notebook version, StratGroupKFold's 0.78 appears higher than StratKFold's 0.78, despite both showing the same leaderboard score.</p>\n<p><a href=\"https://www.kaggle.com/janbrederecke\" target=\"_blank\">@janbrederecke</a> thx for the info!</p>",
          "votes": 5,
          "replies": [
            {
              "id": 3221842,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-11T13:50:04.193000",
              "content": "<p>That is exactly what I found. Thanks for sharing your results as well! 🙂 </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3222109,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-11T20:20:30.610000",
              "content": "<p>Tried the experiment as well, 0.781 cv (sgkf, group on subject, target on handedness-sex-adult_child concat) and 0.79 lb, same lb as using stratified k fold on gesture (which had 0.8 cv).</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3222129,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-11T21:30:37.167000",
              "content": "<p>So does this prevent overfitting, I am not yet 100% sure…</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3226589,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-17T21:01:42.223000",
              "content": "<p>What do you make of this? Would be interested in you perspective on these results.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227137,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-18T15:12:18.783000",
              "content": "<p>Through further experiments, I found that the CV-LB correlation is not ideal for the group-by-subject method, especially in cases where CV increases but LB decreases. I believe this approach leads to more inconsistent results across folds, which may be reflected in the average CV score. In contrast, a simple stratified split yields a more stable and consistent correlation for me, even though facing some leakage in demo data.</p>",
              "votes": 3,
              "replies": []
            },
            {
              "id": 3227216,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-06-18T17:00:51.383000",
              "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> So you moved away from the gender + handedness + adult_child StratifiedGroupKFold in favor of StratKFold on label? I've noticed that improvements on cv with the gender + handedness + adult_child split haven't translated to lb. </p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227305,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-18T19:47:39.477000",
              "content": "<p>Yes, and I find it quite difficult to draw a conclusion. Even with the same CV split, the LB for a 0.84 CV is higher than that for 0.855.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227306,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-18T19:52:14.903000",
              "content": "<p>I also find it hard, I get very similar LB results regardless of the split method, although regular stratified seems to have a lower std on 5 fold compared to gkfold. There's only 81 subjects so it's bound to be noisy. As to which one to trust over the other, don't really know, I get slightly better results with the regular stratified k fold on gesture.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3227310,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-18T20:01:16.017000",
              "content": "<p>More or less the same here. And I hoped you guys could give me a definitive answer for this 😅. Thanks for sharing, appreciate it! 🙂</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227318,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-06-18T20:37:31.373000",
              "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> <a href=\"https://www.kaggle.com/andreizamfir\" target=\"_blank\">@andreizamfir</a> <a href=\"https://www.kaggle.com/janbrederecke\" target=\"_blank\">@janbrederecke</a> Alright, since we're all active, who wants to share how preprocessing has impacted their results? Scaling by segment, by subject, by similar subjects, and global. </p>\n<p>My best submission was my first. No feature engineering, no scaling, just 1D-CNN action. A little perplexing seeing as research related to this problem always entails some scaling. </p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227328,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-18T20:48:33.010000",
              "content": "<p>I am just using StandardScaler(), but I will try no scaling</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3227360,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-18T22:32:50.173000",
              "content": "<p>I have to admit it is exactly the same for me - even though I tried a lot since then but no kind of preprocessing has helped so far. Would be really interested to find something that helps though!<br>\n<a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> , did you use standard scaler per sequence?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227363,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-18T22:43:11.173000",
              "content": "<p>I do absolutely no scaling at the moment, but it’s something I have on my list to revisit. I remember scaling in the beginning (per sequence), and seeing improvements after removing that and I just kept going as is since then.</p>\n<p>I do, however, employ FE and augmentations, both of which improved score in my case.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227364,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-18T22:47:48.147000",
              "content": "<p><a href=\"https://www.kaggle.com/andreizamfir\" target=\"_blank\">@andreizamfir</a> would you mind sharing if you use 1d cnn here as well?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227365,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-18T22:51:04.173000",
              "content": "<p>Yeah, 1D CNN, arch changes provided biggest boosts in score. Think there’s a lot more to squeeze still</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227645,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-19T06:43:20.633000",
              "content": "<p><a href=\"https://www.kaggle.com/janbrederecke\" target=\"_blank\">@janbrederecke</a> yes, per sequence</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227676,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-19T07:09:25.267000",
              "content": "<p>Thanks guys, appreciate the open discussion in this comp! 😊</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3235166,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-06-28T19:25:08.740000",
              "content": "",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3235382,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-29T06:41:19.467000",
              "content": "<p>no gesture used. In my view, every subjects have all gestures, therefore it is not informative to group the subjects using gestures</p>",
              "votes": 1,
              "replies": []
            }
          ]
        },
        {
          "id": 3227784,
          "author_name": "Ryenhails",
          "author_url": "",
          "post_date": "2025-06-19T09:17:38.260000",
          "content": "<p>Update: <br>\nremoved StandardScaler() applied per sequence and the standardization of the demo column. This led to a improvement in LB score, but also introduced more cv-lb gap / data leakage when using a simple stratified split. <br>\nCV: 0.8022 → 0.8481 <br>\nLB: 0.79 → 0.80</p>",
          "votes": 2,
          "replies": [
            {
              "id": 3227786,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-19T09:18:50.977000",
              "content": "<p>Why do you call it data leakage? I am also wondering what that cv lb gap is exactly that is why I ask. 😊</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227793,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-19T09:30:16.083000",
              "content": "<p>I believe the gap between the CV and LB scores can largely be attributed to two factors. First, external factors: differences in distributions between the test and train sets. Second, internal factors: the cross-validation split strategy within the train set, which can lead to data leakage from the training data into the validation folds.</p>\n<p>In my view, removing the StandardScaler() has minimal impact since it only scales information intrinsic to each sequence. In contrast, omitting the standardization of the demo column may lead to more severe data leakage. The scale of the demo column is significantly larger than that of the other features, which might cause the model to over-rely on these features for its predictions. Additionally, using a simple stratified split causes the same demo values to appear in both the training and validation sets, potentially widening the CV-LB gap.</p>\n<p>That said, this is merely my intuition—reality might prove otherwise.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227796,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-19T09:40:45.453000",
              "content": "<p>The data leakage when using demo data between training and validation is a very valid point. That’s where group k folding could avoid this sort of leakage.</p>\n<p>I think it’s part of the reason that you see this gap widening. </p>\n<p>I am not using demo data at all and my cv-lb is 0.84-0.82 on IMU data; this gap was consistent for me for the past 2 weeks.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3227804,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-19T09:53:07.947000",
              "content": "<p>Thanks, I will try no demo training</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227815,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-19T10:07:49.853000",
              "content": "<p>just for my understanding: no demo means, you just do group k fold using subject as the group and gesture as y? And for deeper understanding, the idea is that stratification introduces a kind of leakage into all the folds? Thanks so much for sharing I think this is a great perspective, will also try without if this is what you meant. </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227830,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-19T10:26:17.490000",
              "content": "<p>So when I say I don’t use demo I meant I don’t use any sort of features based on the subject (age, sex, height etc) in training. I also regularly use StratifiedKFold based on gesture. </p>\n<p>Using StratifiedGroupKFold based on the subject as group and stratification based on some demographics still wouldn’t use demo features for training, it would just be a means of splitting data.</p>\n<p>The moment you do regular stratified k fold but also use demo features in training (feed them to whatever model you use), the situation described above by Ryenhails, you will have some of the same features in validation (as you’ll get a mixture of gestures from a particular subject in train and in val, and demo features are static and stricly subject dependent), hence the data leakage.</p>\n<p>(Above message is based on my understanding/intuition of what Ryenhails described, might be wrong!)</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227832,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-19T10:28:20.017000",
              "content": "<p>I believe that if the model does not use the demo columns, there would be no explicit leakage—since the same demo row won’t be split between training and validation, regardless of whether you use GroupKFold, stratified GroupKFold, or a simple stratified split. However, some latent leakage may still occur with a stratified split, as the same subjects could end up in both the training and validation sets.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227835,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-19T10:35:09.710000",
              "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> wouldn't this be prevented by using stratifiedgroupkfold and group by subject? </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227837,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-19T10:36:46.683000",
              "content": "<p>That's exactly what I am thinking as well, that inherent leakage is there with regular stratified split, but it's more pronounced when you also use demographic features in training. Group K folding on subjects eliminates those particular leakage sources, although it causes noisy performance due to small number of subjects. Saving grace is the fact that each subject performed all gestures so you can maintain stratification of gestures with it.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3227838,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-19T10:43:42.273000",
              "content": "<p>So what is our collective verdict rn? Use stratified group k fold with subject as the group and stratify by gesture? Or stratify by gesture + some other demographics? I have not used demographics in training so far and stratify by gesture and multiple demographics for now but experience the cv = lb + 0.01 discrepancy consistently.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3227974,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-06-19T14:05:51.500000",
              "content": "",
              "votes": -1,
              "replies": []
            },
            {
              "id": 3228012,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-19T15:07:08.740000",
              "content": "<p>for me, it is no demo, no group and use simple gesture stratify</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3228040,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-06-19T15:38:53.327000",
              "content": "<p>Appreciate it. Congrats on moving to 2nd! </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3230146,
              "author_name": "minerppdy",
              "author_url": "",
              "post_date": "2025-06-22T14:22:51.873000",
              "content": "<p>impressive score! Was this achieved using just 7 IMU columns? Were orientation behavior data used as supplements?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3230157,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-22T14:35:26.337000",
              "content": "<p>just 7 imu cols</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3235535,
              "author_name": "Rustam Bazarbayev",
              "author_url": "",
              "post_date": "2025-06-29T11:08:25.223000",
              "content": "<p>Really? I cannot improve my LB score beyond 0.7</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3239480,
              "author_name": "Maxim K. Surkov",
              "author_url": "",
              "post_date": "2025-07-02T21:15:03.800000",
              "content": "<p>Wow, I am really excited to know, how many weights in your neural network? 😱</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3240189,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-07-03T14:51:17.233000",
              "content": "<p>model file ~100MB</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3241182,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-04T16:05:12.260000",
              "content": "<p>Is this per fold on IMU only?  </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3241332,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-07-04T18:03:58.887000",
              "content": "<p>yes per fold</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3241413,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-04T19:16:14.520000",
              "content": "<p>When you said iterate the arch, you meant it 😂That's over 3x my current fold sizes.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3241473,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-07-04T20:23:07.613000",
              "content": "<p>🤣🤣 I was just adding more and more blocks, but now it doesn’t seem to be improved anymore.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3241540,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-04T21:30:50.263000",
              "content": "<p><img src=\"https://tenor.com/view/star-wars-more-gif-21856591\" alt=\"MORE!\"></p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3241548,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-04T21:38:58.483000",
              "content": "<p>I’m increasing my model size and seeing CV improvements but I can’t see myself reaching 100MB lmao. Bro is out here 10 blocks deep.. I’ve got four blocks - any more and I overfit, even with lr warmup. </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3251244,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-07-20T08:46:42.057000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3229540,
          "author_name": "menglei",
          "author_url": "",
          "post_date": "2025-06-21T16:27:13.957000",
          "content": "<p>Very impressive, I'm wondering if you are also using the demo features for this, or just 1d cnn for the imu data?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3229561,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-21T17:13:00.707000",
              "content": "<p>no demo, only 1dcnn</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3231367,
          "author_name": "Timmy Juicehouse",
          "author_url": "",
          "post_date": "2025-06-24T08:32:52.463000",
          "content": "<p><a href=\"https://www.kaggle.com/ryenhails\" target=\"_blank\">@ryenhails</a> hello, Ryenhails, Amazing lb. Is 0.78 still your highest score with only IMU now? If it's convenient.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3231433,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-24T10:34:19.127000",
              "content": "<p>Hi, now my best IMU only cv is :0.8036(group by subject)/0.8522(without group), lb:0.81</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3231455,
              "author_name": "Timmy Juicehouse",
              "author_url": "",
              "post_date": "2025-06-24T11:34:51.557000",
              "content": "<p>Thank you, it confirms some of my ideas.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3231642,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-06-24T17:12:40.973000",
              "content": "<p>I’m very curious how you’ve managed to separate eyebrow and eyelash hair pulls. I’ve got some ideas I’ll experiment with later… once I figure it out I’ll be right there with ya</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3231670,
              "author_name": "Ryenhails",
              "author_url": "",
              "post_date": "2025-06-24T18:03:04.553000",
              "content": "<p>I have not check the specific class yet, just iterate the arch</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3250014,
      "author_name": "I2nfinit3y",
      "author_url": "",
      "post_date": "2025-07-17T14:41:00.573000",
      "content": "<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.8132</td>\n<td>0.82</td>\n</tr>\n<tr>\n<td>0.7968</td>\n<td>0.81</td>\n</tr>\n</tbody>\n</table>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3225837,
      "author_name": "moth",
      "author_url": "",
      "post_date": "2025-06-16T22:56:27.597000",
      "content": "<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.744</td>\n<td>0.75</td>\n</tr>\n<tr>\n<td>0.740</td>\n<td>0.76</td>\n</tr>\n<tr>\n<td>0.748</td>\n<td>0.75</td>\n</tr>\n<tr>\n<td>0.764</td>\n<td>0.77</td>\n</tr>\n</tbody>\n</table>\n<ul>\n<li>2nd 9 classes</li>\n<li>rest 18 classes</li>\n</ul>\n<p>CV: StratifiedGroupKFold -&gt; group=subject, target=gesture</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3225814,
      "author_name": "lightsource<3",
      "author_url": "",
      "post_date": "2025-06-16T21:46:21.687000",
      "content": "<p>cv - 0.7688<br>\nlb - 0.78 (mid)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3249843,
          "author_name": "lightsource<3",
          "author_url": "",
          "post_date": "2025-07-17T08:06:32.273000",
          "content": "<p>sgkf(group=subject, y=gesture, n_folds=5)</p>\n<p>cv - 0.826<br>\nlb - 0.82</p>",
          "votes": 1,
          "replies": [
            {
              "id": 3249852,
              "author_name": "PatrickAIForFun",
              "author_url": "",
              "post_date": "2025-07-17T08:23:48.447000",
              "content": "<p>Still IMU only? That's impressive</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3249856,
              "author_name": "lightsource<3",
              "author_url": "",
              "post_date": "2025-07-17T08:34:15.993000",
              "content": "<p>yes, imu only</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3249982,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-17T13:24:38.250000",
              "content": "<p>You should be able to get top 30 right now with that. Features or architecture get you there? Or something else? </p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3249984,
              "author_name": "lightsource<3",
              "author_url": "",
              "post_date": "2025-07-17T13:33:16.450000",
              "content": "<p>both of them</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3249995,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-17T13:59:28.603000",
              "content": "<p>I'm curious what features beyond public notebooks you've tried that were actually helpful. I've tried csums, diffs, longer ranged diffs, shifts, diffs from sequence median, wavelets, and some others. I even tried the 2d wavelets/spectograms as you did.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3249997,
              "author_name": "lightsource<3",
              "author_url": "",
              "post_date": "2025-07-17T14:02:41.837000",
              "content": "<p>nothing other than what you listed</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3249998,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2025-07-17T14:14:11.957000",
              "content": "<p>And some of these were helpful for you? I've found my best set of features is even smaller than what most public notebooks have.. hmm</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3261867,
              "author_name": "Mahog",
              "author_url": "",
              "post_date": "2025-08-02T11:56:23.777000",
              "content": "<p>Did you do arch tuning? If you did then how much did it improve CV/LB?</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3222454,
      "author_name": "Angle",
      "author_url": "",
      "post_date": "2025-06-12T07:53:11.023000",
      "content": "<p>using StratifiedGroupKFold (5fold, y is gesture, groups is subject)<br>\nOnly IMU:<br>\nCV: 0.795<br>\nLB: 0.75</p>\n<p>Amazing results…</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3222022,
      "author_name": "Natsume",
      "author_url": "",
      "post_date": "2025-06-11T18:04:41.997000",
      "content": "<p>cv-0.7583<br>\nlb-0.78<br>\nonly IMU</p>\n<p>update: cv-0.7714,lb-0.79</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3222534,
          "author_name": "Rustam Bazarbayev",
          "author_url": "",
          "post_date": "2025-06-12T08:43:04.927000",
          "content": "<p>What model did you use?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3222594,
              "author_name": "Natsume",
              "author_url": "",
              "post_date": "2025-06-12T09:34:58.760000",
              "content": "<p>just 1d-cnn</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3226584,
              "author_name": "Jan Brederecke",
              "author_url": "",
              "post_date": "2025-06-17T20:59:35.180000",
              "content": "<p>Can you share how you create folds? Do you do stratification and if so only by the gestures or something else? I observe a similar cv / lb relationship using stratification.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3221623,
      "author_name": "Karearea",
      "author_url": "",
      "post_date": "2025-06-11T09:10:59.717000",
      "content": "<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.762</td>\n<td>0.75</td>\n</tr>\n</tbody>\n</table>\n<p>5-fold catboost model<br>\nOut of ideas for now, time to look at the other sensors<br>\nAfter reading the comments here, might need to experiment with a different stratification for my cv too</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3218831,
      "author_name": "Ravi Ramakrishnan",
      "author_url": "",
      "post_date": "2025-06-06T19:03:43.033000",
      "content": "<p><a href=\"https://www.kaggle.com/gerwynng\" target=\"_blank\">@gerwynng</a> just starting out<br>\n<a href=\"https://www.kaggle.com/code/ravi20076/cmi2025-public-baseline-v3\" target=\"_blank\">Single catboost</a> CV 0.6904, LB = 0.69</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3217902,
      "author_name": "Carlos Pérez Ricardo",
      "author_url": "",
      "post_date": "2025-06-05T13:59:34.467000",
      "content": "<p>CV = 0.723 (cv scores vary between 0.69-0.76 on 5 folds)<br>\nLB = 0.73 <br>\nUsing DL model with 1d CNN + LSTM. <br>\nProbably with an ensemble we can improve scores but now I will move to other sources of data.</p>\n<p>Edit:<br>\nLB = 0.75</p>\n<p>Edit 2:<br>\nSplit Stratified Group KFold (group=subject, target=sequence_type)<br>\nCV = 0.806, LB = 0.80 (mid)</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3217887,
      "author_name": "Andrei Zamfir",
      "author_url": "",
      "post_date": "2025-06-05T13:39:07.683000",
      "content": "<p>IMU Only, 2  1D CNN layers  1 dense layer starter:</p>\n<pre><code> of validation scores:\n   mean   : . ± .\n   binary : . ± .\n   macro  : . ± .\n</code></pre>\n<pre><code> : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n : F1 mean=., F1 binary=., F1 macro=.\n</code></pre>\n<p>L.E.: LB: 0.73</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3217890,
          "author_name": "Ryenhails",
          "author_url": "",
          "post_date": "2025-06-05T13:42:04.347000",
          "content": "<p>Impressive! what is your lb according to this cv</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3217896,
              "author_name": "Andrei Zamfir",
              "author_url": "",
              "post_date": "2025-06-05T13:50:38.983000",
              "content": "<p>Haven't sent it out yet, will probably do today and come back with an update.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3217700,
      "author_name": "Arunodhayan",
      "author_url": "",
      "post_date": "2025-06-05T10:24:08.920000",
      "content": "<p>only IMU 0.68</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3217201,
      "author_name": "Gunes Evitan",
      "author_url": "",
      "post_date": "2025-06-04T16:48:09.417000",
      "content": "<p>I did very few experiments so far and my current imu only scores are like this right now. I'll update this when I have some improvements.</p>\n<pre><code>  : ,\n  : ,\n  : \n</code></pre>",
      "votes": 1,
      "replies": [
        {
          "id": 3217271,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-04T18:46:00.003000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3217360,
          "author_name": "moth",
          "author_url": "",
          "post_date": "2025-06-04T23:58:13.763000",
          "content": "<p>Im getting the same binary score, macro still around 0.36, average ~0.66. I dont know why Im having trouble breaking this threshold. Validation is stratified group kfold with 5 folds, using subject as group, no demographics whatsoever</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3218051,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-06-05T17:50:03.517000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3217777,
          "author_name": "Gerwyn",
          "author_url": "",
          "post_date": "2025-06-05T11:50:51.087000",
          "content": "<p>mine just hovers around my best cv:</p>\n<pre><code>\n     \n     \n     \n\n</code></pre>\n<p>usually <code>f1_binary</code> ~ 0.960, <code>f1_macro</code> ~0.49X</p>\n<p>would like to say that i have exhausted IMU so i can move on haha</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 3217166,
      "author_name": "Ryenhails",
      "author_url": "",
      "post_date": "2025-06-04T15:39:59.367000",
      "content": "<table>\n<thead>\n<tr>\n<th>5folds cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.6980</td>\n<td>0.69</td>\n</tr>\n<tr>\n<td>0.7137</td>\n<td>0.71</td>\n</tr>\n<tr>\n<td>0.7329</td>\n<td>0.73</td>\n</tr>\n</tbody>\n</table>",
      "votes": 1,
      "replies": [
        {
          "id": 3217779,
          "author_name": "Gerwyn",
          "author_url": "",
          "post_date": "2025-06-05T11:52:13.933000",
          "content": "<p>ahh nice, seems to corroborate with my experience, that scores <em>start to</em> plateau around 0.73 with IMU only </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3224688,
      "author_name": "Sinan Calisir",
      "author_url": "",
      "post_date": "2025-06-15T10:27:30.480000",
      "content": "<p>Only IMU</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.728</td>\n<td>0.72</td>\n</tr>\n<tr>\n<td>0.748</td>\n<td>0.74</td>\n</tr>\n</tbody>\n</table>",
      "votes": 2,
      "replies": [
        {
          "id": 3229538,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-21T16:25:14.890000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3218596,
      "author_name": "Mert Bayraktar",
      "author_url": "",
      "post_date": "2025-06-06T12:06:31.020000",
      "content": "<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.724</td>\n<td>0.72</td>\n</tr>\n<tr>\n<td>0.751</td>\n<td>0.74</td>\n</tr>\n</tbody>\n</table>\n<p>only imu and default params for gbdt. still a lot to do.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3221340,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-10T19:40:32.830000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 3237027,
          "author_name": "Mert Bayraktar",
          "author_url": "",
          "post_date": "2025-06-30T20:30:39.867000",
          "content": "<p>Update:</p>\n<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.758</td>\n<td>0.75</td>\n</tr>\n<tr>\n<td>0.763</td>\n<td>0.75</td>\n</tr>\n</tbody>\n</table>\n<p>still only imu and GBDT, it's hard to beat CNN-based models, now I can move on to them.</p>\n<p>Update:</p>\n<p>switched to CNN, initial experiments:</p>\n<table>\n<thead>\n<tr>\n<th>cv</th>\n<th>lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.742</td>\n<td>0.77</td>\n</tr>\n</tbody>\n</table>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3218227,
      "author_name": "Timmy Juicehouse",
      "author_url": "",
      "post_date": "2025-06-06T02:10:20.537000",
      "content": "<p>first trial 5-fold cv: 0.717 lb:0.7</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3220476,
          "author_name": "Timmy Juicehouse",
          "author_url": "",
          "post_date": "2025-06-09T11:41:01.820000",
          "content": "<p>Update today:</p>\n<table>\n<thead>\n<tr>\n<th>Cv</th>\n<th>Lb</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.7085</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7099</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7109</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7121</td>\n<td>0.7</td>\n</tr>\n<tr>\n<td>0.7195</td>\n<td>0.71</td>\n</tr>\n<tr>\n<td>0.7216</td>\n<td>0.72</td>\n</tr>\n</tbody>\n</table>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3217632,
      "author_name": "Ertuğrul Demir",
      "author_url": "",
      "post_date": "2025-06-05T08:22:05.343000",
      "content": "<h4>My First Test with IMU Only</h4>\n<p>Here are the initial results:</p>\n<table>\n<thead>\n<tr>\n<th>5 Fold OOF CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.7052</td>\n<td>0.70</td>\n</tr>\n</tbody>\n</table>\n<p>Seem like a lot to improve upon…</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3218587,
          "author_name": "Ertuğrul Demir",
          "author_url": "",
          "post_date": "2025-06-06T11:50:11.017000",
          "content": "<table>\n<thead>\n<tr>\n<th>5 Fold OOF CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.7458</td>\n<td>0.74</td>\n</tr>\n</tbody>\n</table>\n<p>Still IMU only and no cnn or other seq models. Wonder what's the extend 🤔</p>",
          "votes": 2,
          "replies": [
            {
              "id": 3218592,
              "author_name": "moth",
              "author_url": "",
              "post_date": "2025-06-06T11:53:50.080000",
              "content": "<p>Are these results for GBM models?</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3218594,
              "author_name": "Gerwyn",
              "author_url": "",
              "post_date": "2025-06-06T11:57:42.893000",
              "content": "<p>i havent been able to surpass 0.73X, good to know 0.74 can be done 😃</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3218595,
              "author_name": "Ertuğrul Demir",
              "author_url": "",
              "post_date": "2025-06-06T11:58:24.830000",
              "content": "<p><a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a> </p>\n<p>Yeah, almost with default hparams, not optimized at all</p>",
              "votes": 1,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3248648,
      "author_name": "kmn",
      "author_url": "",
      "post_date": "2025-07-15T00:44:33.510000",
      "content": "<p>cv - 0.7983<br>\nlb - 0.80 (10-fold)</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3243986,
      "author_name": "Ryan Barretto",
      "author_url": "",
      "post_date": "2025-07-07T17:30:17.037000",
      "content": "<p>StratifiedGroupKFold group=subject, y=gesture</p>\n<p>cv: 0.797<br>\nlb: 0.8</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3243928,
      "author_name": "Desolade",
      "author_url": "",
      "post_date": "2025-07-07T16:49:35.173000",
      "content": "<p>1D CNN IMU only    (5 folds group stratified)</p>\n<table>\n<thead>\n<tr>\n<th>CV</th>\n<th>LB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>0.791</td>\n<td>0.78</td>\n</tr>\n<tr>\n<td>0.780</td>\n<td>0.77</td>\n</tr>\n</tbody>\n</table>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3243876,
      "author_name": "Gnanasenthil",
      "author_url": "",
      "post_date": "2025-07-07T15:56:53.857000",
      "content": "<p>Hi,<br>\nI did merge demographics and training data. Initially when I used all sensor data, the model performed very poorly (0.18) but when the IMU data alone was used (still merged with demographics) the accuracy improved (0.52). Still haven't used the test.csv though.</p>\n<p>But how to improve further..</p>\n<p>I have attached my notebook below..</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3242945,
      "author_name": "Kurise",
      "author_url": "",
      "post_date": "2025-07-06T15:10:42.410000",
      "content": "<p>StratifiedGroupKFold, split by subject: cv : 0.7789 LB : 0.79</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3243358,
          "author_name": "Kurise",
          "author_url": "",
          "post_date": "2025-07-07T04:09:15.933000",
          "content": "<p>test1 : CV : 0.7904 LB : 0.81 (low)</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3245809,
              "author_name": "menglei",
              "author_url": "",
              "post_date": "2025-07-09T22:19:24.323000",
              "content": "<p>is that imu only？</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3241015,
      "author_name": "Tom",
      "author_url": "",
      "post_date": "2025-07-04T12:48:02.290000",
      "content": "<p>I'm really curious whether existing a team approaches the information of tof and thm only using imu.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3240995,
      "author_name": "Gerwyn",
      "author_url": "",
      "post_date": "2025-07-04T12:29:50.203000",
      "content": "<p>update: just did my first NN model submission:</p>\n<p>cv - 0.7952, LB - 0.79</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3240924,
      "author_name": "Tomoki Hirose",
      "author_url": "",
      "post_date": "2025-07-04T10:49:30.690000",
      "content": "<p>I am using StratifiedKFold with the variables sequence_id and gesture.<br>\ncv 0.8755 / LB 0.80 🤣</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3240993,
          "author_name": "Jan Brederecke",
          "author_url": "",
          "post_date": "2025-07-04T12:27:48.843000",
          "content": "<p>Wow, what a CV 😅. Anything to share about that? Kudos if that is correct! </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3240759,
      "author_name": "PiGram",
      "author_url": "",
      "post_date": "2025-07-04T08:01:13.083000",
      "content": "<p>Only IMU - CV = 0.83; LB=0.81</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3241422,
          "author_name": "Vishal Kishore",
          "author_url": "",
          "post_date": "2025-07-04T19:27:20.247000",
          "content": "<p>What kind of model are you missing? If possible. I am stuck at 0.76 with 1d cnn imu only model</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3239538,
      "author_name": "menglei",
      "author_url": "",
      "post_date": "2025-07-02T23:25:53.377000",
      "content": "<p>CV:0.82, LB:0.78 </p>",
      "votes": 0,
      "replies": [
        {
          "id": 3246563,
          "author_name": "menglei",
          "author_url": "",
          "post_date": "2025-07-11T06:14:46.243000",
          "content": "<p>update cv: 0.79, LB 0.79</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3233103,
      "author_name": "Mahdi Ravaghi",
      "author_url": "",
      "post_date": "2025-06-26T13:35:10.560000",
      "content": "<table>\n<thead>\n<tr>\n<th><strong>Date</strong></th>\n<th><strong>5 Fold CV</strong></th>\n<th><strong>Public LB</strong></th>\n<th><strong>CV Strategy</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>26. June</td>\n<td>0.759124</td>\n<td>0.74</td>\n<td>StratifiedGroupKFold, groups=subject</td>\n</tr>\n</tbody>\n</table>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3232810,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-26T08:37:09.647000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 3233030,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-26T12:33:06.850000",
          "content": "",
          "votes": 0,
          "replies": [
            {
              "id": 3233949,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-06-27T11:15:03.553000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3244292,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-07-08T04:14:07.023000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3230797,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-23T12:56:51.077000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 3230827,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-23T13:45:48.757000",
          "content": "",
          "votes": 0,
          "replies": [
            {
              "id": 3230849,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-06-23T14:30:18.377000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3246558,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-07-11T06:09:17.067000",
          "content": "",
          "votes": -6,
          "replies": []
        }
      ]
    },
    {
      "id": 3230252,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-22T16:10:18.350000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3228985,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-20T19:30:40.470000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 3246565,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-07-11T06:15:57.313000",
          "content": "",
          "votes": 0,
          "replies": [
            {
              "id": 3246736,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-07-11T12:34:39.763000",
              "content": "",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3246758,
              "author_name": "",
              "author_url": "",
              "post_date": "2025-07-11T13:40:22.657000",
              "content": "",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3227729,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-19T08:21:25.057000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 3227780,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-19T09:09:26.310000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3225581,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-16T15:46:14.017000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3225197,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-16T05:14:34.713000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3218385,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-06T05:57:26.213000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3217520,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-05T05:16:05.563000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3216977,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-04T11:20:05.540000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 3216992,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-04T11:41:09.630000",
          "content": "",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 3228874,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-20T16:03:22.177000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3228550,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-20T09:23:18.243000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3216928": "Mine currently:\n\n|  cv | lb |\n| --- | --- |\n| 0.72109 | 0.73  |\n| 0.72974 | 0.72 |\n| 0.74759 | 0.74 |\n|0.76612 | 0.75 |\n\n- CV using group on subject - stratified - 5 folds. (OOF vs CV-Average-Scores across 5-folds are very similar)\n- LB is 5 folds predictions into majority voting.\n\n\nI am curious what are your experience with IMU-only performance\n\n\nedit 1: LB 0.74 \nedit 2: LB 0.75\n",
    "3271800": "CV: 0.845\nLB: 0.831",
    "3218857": "update: \nIMU only, 1d cnn\ncv: 0.7782 lb:0.76",
    "3250014": "| CV | LB |\n| --- | --- |\n| 0.8132| 0.82 |\n| 0.7968 | 0.81 |\n",
    "3225837": "| CV | LB |\n| --- | --- |\n|0.744  |0.75  | \n|0.740  |0.76  |\n|0.748  |0.75  |\n|0.764  |0.77  |\n\n- 2nd 9 classes\n- rest 18 classes\n\nCV: StratifiedGroupKFold -> group=subject, target=gesture",
    "3225814": "cv - 0.7688\nlb - 0.78 (mid)",
    "3222454": "using StratifiedGroupKFold (5fold, y is gesture, groups is subject)\nOnly IMU:\nCV: 0.795\nLB: 0.75\n\nAmazing results...",
    "3222022": "cv-0.7583\nlb-0.78\nonly IMU\n\nupdate: cv-0.7714,lb-0.79",
    "3221623": "| CV | LB |\n| --- | --- |\n| 0.762 | 0.75 |\n\n5-fold catboost model\nOut of ideas for now, time to look at the other sensors\nAfter reading the comments here, might need to experiment with a different stratification for my cv too",
    "3218831": "@gerwynng just starting out\n[Single catboost](https://www.kaggle.com/code/ravi20076/cmi2025-public-baseline-v3) CV 0.6904, LB = 0.69",
    "3217902": "CV = 0.723 (cv scores vary between 0.69-0.76 on 5 folds)\nLB = 0.73 \nUsing DL model with 1d CNN + LSTM. \nProbably with an ensemble we can improve scores but now I will move to other sources of data.\n\nEdit:\nLB = 0.75\n\nEdit 2:\nSplit Stratified Group KFold (group=subject, target=sequence_type)\nCV = 0.806, LB = 0.80 (mid)",
    "3217887": "IMU Only, 2  1D CNN layers  1 dense layer starter:\n\n```\nMean of validation scores:\n  F1 mean   : 0.7404 ± 0.0072\n  F1 binary : 0.9611 ± 0.0018\n  F1 macro  : 0.5196 ± 0.0135\n```\n\n```\nFold 0: F1 mean=0.7450, F1 binary=0.9642, F1 macro=0.5258\nFold 1: F1 mean=0.7409, F1 binary=0.9598, F1 macro=0.5220\nFold 2: F1 mean=0.7287, F1 binary=0.9590, F1 macro=0.4985\nFold 3: F1 mean=0.7371, F1 binary=0.9615, F1 macro=0.5127\nFold 4: F1 mean=0.7501, F1 binary=0.9611, F1 macro=0.5391\n```\n\nL.E.: LB: 0.73",
    "3217700": "only IMU 0.68",
    "3217201": "I did very few experiments so far and my current imu only scores are like this right now. I'll update this when I have some improvements.\n\n```python\n  \"f1_binary\": 0.9584695250913285,\n  \"f1_macro\": 0.4780790342101205,\n  \"f1_average\": 0.7182742796507245\n```",
    "3217166": "|5folds cv | lb |\n| --- | --- |\n| 0.6980 | 0.69 |\n| 0.7137 | 0.71 |\n| 0.7329 | 0.73 |\n",
    "3224688": "Only IMU\n\n| CV    | LB   |\n|-------|------|\n| 0.728 | 0.72 |\n| 0.748 | 0.74 |",
    "3218596": "| cv | lb |\n| --- | --- |\n| 0.724 | 0.72 |\n| 0.751 | 0.74 |\n\n\nonly imu and default params for gbdt. still a lot to do.",
    "3218227": "first trial 5-fold cv: 0.717 lb:0.7",
    "3217632": "#### My First Test with IMU Only\n\nHere are the initial results:\n\n| 5 Fold OOF CV     | LB   |\n|--------|------|\n| 0.7052 | 0.70 |\n\nSeem like a lot to improve upon...",
    "3248648": "cv - 0.7983\nlb - 0.80 (10-fold)",
    "3243986": "StratifiedGroupKFold group=subject, y=gesture\n\ncv: 0.797\nlb: 0.8",
    "3243928": "1D CNN IMU only    (5 folds group stratified)\n| CV | LB |\n| --- | --- |\n| 0.791 |0.78 |\n| 0.780|0.77 |\n",
    "3243876": "Hi,\nI did merge demographics and training data. Initially when I used all sensor data, the model performed very poorly (0.18) but when the IMU data alone was used (still merged with demographics) the accuracy improved (0.52). Still haven't used the test.csv though.\n\nBut how to improve further..\n\nI have attached my notebook below..",
    "3242945": "StratifiedGroupKFold, split by subject: cv : 0.7789 LB : 0.79",
    "3241015": "I'm really curious whether existing a team approaches the information of tof and thm only using imu.",
    "3240995": "update: just did my first NN model submission:\n\ncv - 0.7952, LB - 0.79",
    "3240924": "I am using StratifiedKFold with the variables sequence_id and gesture.\ncv 0.8755 / LB 0.80 🤣",
    "3240759": "Only IMU - CV = 0.83; LB=0.81",
    "3239538": "CV:0.82, LB:0.78 ",
    "3233103": "| **Date** | **5 Fold CV** | **Public LB** |            **CV Strategy**           |\n|:--------:|:-------------:|:-------------:|:------------------------------------:|\n| 26. June |      0.759124     |      0.74     | StratifiedGroupKFold, groups=subject |",
    "3232810": "IMU + THM + TOF, StratifiedGroupKFold:\n- CV: 0.8027. Public leader board: 0.81.\n- CV: 0.8840. Public leader board: 0.81.\n🤔😅",
    "3230797": "IMU only 5 folds group stratified (group=subject, y=sequence_type), no demographics.\n\n| CV | LB (single fold) | LB (5 fold avg)|\n| --- | --- |---|\n| 0.756 | 0.74 (mid) | 0.75 (mid-high)|\n| 0.783 | NaN | 0.78 (low)|",
    "3230252": "CV 0.728 LB 0.74",
    "3228985": "cv: 0.7915, lb: 0.78\n\nupdate: \ncv: 0.8121 lb: 0.8\n\nupdate:\ncv: 0.8209, lb: 0.81\n\nlb: 0.82 ",
    "3227729": "IMU ONLY:\nCV:0.7738  LB:0.76\nIMU+TOF:\nCV:0.85  LB:0.79",
    "3225581": "GroupKfold stratified based on subject and gesture\nCV = 0.6742\nLB = 0.68",
    "3225197": "Update:\ncv - 0.76612\nlb - 0.75\n\n\nI have several improvements (different enough experiments) that improved CV but did not translate into improvements on LB. I would think it's quite easy to overfit to a few sequence_ids, once performance is good enough. ",
    "3218385": "`accuracy: 0.9753\n f1_macro: 0.9762 \n loss: 0.4299`\nLB: 0.69",
    "3217520": "After doing groupby are you removing duplicates? aren't that reducing the data shape significantly?",
    "3216977": "May I ask if your score is using the raw sequences or did you apply any preprocessing besides padding, scaling, trimming?",
    "3228874": "",
    "3228550": "@gerwynng IMU data means you are not considering **\"tof_[1-5]_v[0-63]\"** columns ?"
  }
}