I am using LightGBM and CatBoost, and my highest score with this ensemble technique is 0.463.
\nCan you share your model architecture, CV, and LB scores?
\nMy rough estimate at the start of this competition is that we will achieve at least 0.6 within a month, and eventually reach around 0.7.
\nBut, this is computationally very expensive, and we can't add too many features or use heavy model architectures with Kaggle GPUs.
\nPS: One month is left, keep experimenting. Let's see who will achieve the best score.
", "messages": [ { "id": "3328198", "postDate": "11/15/2025 17:36:41", "content": "I am using LightGBM and CatBoost, and my highest score with this ensemble technique is 0.463.
\nCan you share your model architecture, CV, and LB scores?
\nMy rough estimate at the start of this competition is that we will achieve at least 0.6 within a month, and eventually reach around 0.7.
\nBut, this is computationally very expensive, and we can't add too many features or use heavy model architectures with Kaggle GPUs.
\nPS: One month is left, keep experimenting. Let's see who will achieve the best score.
", "rawMarkdown": "I am using LightGBM and CatBoost, and my highest score with this ensemble technique is 0.463.\n\nCan you share your model architecture, CV, and LB scores?\n\nMy rough estimate at the start of this competition is that we will achieve at least 0.6 within a month, and eventually reach around 0.7.\n\nBut, this is computationally very expensive, and we can't add too many features or use heavy model architectures with Kaggle GPUs.\n\nPS: One month is left, keep experimenting. Let's see who will achieve the best score.", "votes": null }, { "id": "3348854", "postDate": "11/26/2025 09:00:19", "content": "I have a CNN performing over 0.434 locally, but I don't know its LB score. (I didnt submit it yet since it is not competitive)
\nI guess it is relatively difficult to make NNs work in this competition due to the sparsity of some labels.
", "rawMarkdown": "I have a CNN performing over 0.434 locally, but I don't know its LB score. (I didnt submit it yet since it is not competitive)\n\nI guess it is relatively difficult to make NNs work in this competition due to the sparsity of some labels.", "votes": null }, { "id": "3349106", "postDate": "11/26/2025 12:53:31", "content": "I tried the GRU model but abandoned it because the samples were too sparse. \nHowever, the main issue was the extremely high feature dimension, making it too costly and less efficient than tree models.
", "rawMarkdown": "I tried the GRU model but abandoned it because the samples were too sparse. \nHowever, the main issue was the extremely high feature dimension, making it too costly and less efficient than tree models.", "votes": null }, { "id": "3351146", "postDate": "11/28/2025 07:20:06", "content": "I tried using 1D TCN, but the score was low and it was very time-consuming, despite my efforts to improve its performance.
", "rawMarkdown": "I tried using 1D TCN, but the score was low and it was very time-consuming, despite my efforts to improve its performance.", "votes": null }, { "id": "3361815", "postDate": "12/03/2025 20:47:06", "content": "I'm currently working with CNN and it can reach 0.422 on LB. Actually the model is quite lightweight if you only focus on the bodyparts that are used the most (I only use 7)
", "rawMarkdown": "I'm currently working with CNN and it can reach 0.422 on LB. Actually the model is quite lightweight if you only focus on the bodyparts that are used the most (I only use 7)", "votes": null }, { "id": "3361990", "postDate": "12/04/2025 03:33:12", "content": "It is weird that nn underperform trees. Maybe all the top ranks are secretly using NN only?
", "rawMarkdown": "It is weird that nn underperform trees. Maybe all the top ranks are secretly using NN only?", "votes": null }, { "id": "3364431", "postDate": "12/06/2025 13:43:18", "content": "Your meta-awareness of Kaggle competitions is truly outstanding.
\nThe accuracy of your predictions seems to be extremely high.
Update: A simple 1D CNN can reach a LB score of 0.457 so it’s definitely worth exploring
", "rawMarkdown": "Update: A simple 1D CNN can reach a LB score of 0.457 so it’s definitely worth exploring", "votes": null }, { "id": "3367717", "postDate": "12/08/2025 17:10:16", "content": "Is this with post-processing?
", "rawMarkdown": "Is this with post-processing?", "votes": null }, { "id": "3368009", "postDate": "12/08/2025 22:11:02", "content": "Not much, I just ignore labels that are not long enough (in my case 8 frames)
", "rawMarkdown": "Not much, I just ignore labels that are not long enough (in my case 8 frames)", "votes": null }, { "id": "3368105", "postDate": "12/09/2025 01:26:46", "content": "What is the training time for your 1d-CNN? And are you still using only 7 bodyparts?
", "rawMarkdown": "What is the training time for your 1d-CNN? And are you still using only 7 bodyparts?", "votes": null }, { "id": "3368360", "postDate": "12/09/2025 07:05:03", "content": "1 hour for training, 30min for inference. Yes, only 7 bodyparts
", "rawMarkdown": "1 hour for training, 30min for inference. Yes, only 7 bodyparts", "votes": null }, { "id": "3369555", "postDate": "12/10/2025 06:16:35", "content": "Hello, may I ask whether you train a separate CNN for each action, or input all the data to directly predict all actions at once?
", "rawMarkdown": "Hello, may I ask whether you train a separate CNN for each action, or input all the data to directly predict all actions at once?", "votes": null }, { "id": "3369590", "postDate": "12/10/2025 07:03:47", "content": "It's one CNN for all the actions at the same time
", "rawMarkdown": "It's one CNN for all the actions at the same time", "votes": null }, { "id": "3369609", "postDate": "12/10/2025 07:20:56", "content": "Thank you for your reply. However, loading the entire dataset at once for processing may not be feasible on Kaggle’s free computing resources—you’re likely using your own local or cloud-based hardware, right?
", "rawMarkdown": "Thank you for your reply. However, loading the entire dataset at once for processing may not be feasible on Kaggle’s free computing resources—you’re likely using your own local or cloud-based hardware, right?", "votes": null }, { "id": "3369673", "postDate": "12/10/2025 08:24:53", "content": "I am using my own hardware indeed. But it should also be possible on Kaggle by doing a first notebook that will preprocess the sequences and then use it as an input. Also, you don't have to train your model in the same notebook as the inference one, so you can train your model first and then use it in the inference notebook.
", "rawMarkdown": "I am using my own hardware indeed. But it should also be possible on Kaggle by doing a first notebook that will preprocess the sequences and then use it as an input. Also, you don't have to train your model in the same notebook as the inference one, so you can train your model first and then use it in the inference notebook.", "votes": null }, { "id": "3369827", "postDate": "12/10/2025 11:32:02", "content": "Tried a TCN plus ST GCN got 0.022 on LB
", "rawMarkdown": "Tried a TCN plus ST GCN got 0.022 on LB", "votes": null }, { "id": "3370795", "postDate": "12/11/2025 05:34:55", "content": "Hi, for data that has fewer than 7 body parts or doesn't match the 7 parts you selected, do you just discard it?
", "rawMarkdown": "Hi, for data that has fewer than 7 body parts or doesn't match the 7 parts you selected, do you just discard it?", "votes": null }, { "id": "3371719", "postDate": "12/11/2025 16:06:31", "content": "No I keep everything and just replace those missing values by (0,0) and the model simply learns to ignore them. From my experiments, too much preprocessing will make the model perform worse.
", "rawMarkdown": "No I keep everything and just replace those missing values by (0,0) and the model simply learns to ignore them. From my experiments, too much preprocessing will make the model perform worse.", "votes": null } ], "comments": [ { "id": 3348854, "author_name": "aerdem4", "author_url": "", "post_date": "11/26/2025 09:00:19", "content": "I have a CNN performing over 0.434 locally, but I don't know its LB score. (I didnt submit it yet since it is not competitive)
\nI guess it is relatively difficult to make NNs work in this competition due to the sparsity of some labels.
", "votes": null, "replies": [] }, { "id": 3349106, "author_name": "suxingstar", "author_url": "", "post_date": "11/26/2025 12:53:31", "content": "I tried the GRU model but abandoned it because the samples were too sparse. \nHowever, the main issue was the extremely high feature dimension, making it too costly and less efficient than tree models.
", "votes": null, "replies": [] }, { "id": 3351146, "author_name": "bioconvolutionyt", "author_url": "", "post_date": "11/28/2025 07:20:06", "content": "I tried using 1D TCN, but the score was low and it was very time-consuming, despite my efforts to improve its performance.
", "votes": null, "replies": [ { "id": 3369827, "author_name": "lukorito", "author_url": "", "post_date": "12/10/2025 11:32:02", "content": "Tried a TCN plus ST GCN got 0.022 on LB
", "votes": null, "replies": [] } ] }, { "id": 3361815, "author_name": "antoinemasq", "author_url": "", "post_date": "12/03/2025 20:47:06", "content": "I'm currently working with CNN and it can reach 0.422 on LB. Actually the model is quite lightweight if you only focus on the bodyparts that are used the most (I only use 7)
", "votes": null, "replies": [ { "id": 3367287, "author_name": "antoinemasq", "author_url": "", "post_date": "12/08/2025 10:24:35", "content": "Update: A simple 1D CNN can reach a LB score of 0.457 so it’s definitely worth exploring
", "votes": null, "replies": [ { "id": 3367717, "author_name": "xiaoz259", "author_url": "", "post_date": "12/08/2025 17:10:16", "content": "Is this with post-processing?
", "votes": null, "replies": [ { "id": 3368009, "author_name": "antoinemasq", "author_url": "", "post_date": "12/08/2025 22:11:02", "content": "Not much, I just ignore labels that are not long enough (in my case 8 frames)
", "votes": null, "replies": [] } ] }, { "id": 3368105, "author_name": "mahoganybuttstrings", "author_url": "", "post_date": "12/09/2025 01:26:46", "content": "What is the training time for your 1d-CNN? And are you still using only 7 bodyparts?
", "votes": null, "replies": [ { "id": 3368360, "author_name": "antoinemasq", "author_url": "", "post_date": "12/09/2025 07:05:03", "content": "1 hour for training, 30min for inference. Yes, only 7 bodyparts
", "votes": null, "replies": [ { "id": 3369555, "author_name": "yaokaixinya123", "author_url": "", "post_date": "12/10/2025 06:16:35", "content": "Hello, may I ask whether you train a separate CNN for each action, or input all the data to directly predict all actions at once?
", "votes": null, "replies": [ { "id": 3369590, "author_name": "antoinemasq", "author_url": "", "post_date": "12/10/2025 07:03:47", "content": "It's one CNN for all the actions at the same time
", "votes": null, "replies": [ { "id": 3369609, "author_name": "yaokaixinya123", "author_url": "", "post_date": "12/10/2025 07:20:56", "content": "Thank you for your reply. However, loading the entire dataset at once for processing may not be feasible on Kaggle’s free computing resources—you’re likely using your own local or cloud-based hardware, right?
", "votes": null, "replies": [ { "id": 3369673, "author_name": "antoinemasq", "author_url": "", "post_date": "12/10/2025 08:24:53", "content": "I am using my own hardware indeed. But it should also be possible on Kaggle by doing a first notebook that will preprocess the sequences and then use it as an input. Also, you don't have to train your model in the same notebook as the inference one, so you can train your model first and then use it in the inference notebook.
", "votes": null, "replies": [ { "id": 3370795, "author_name": "yaokaixinya123", "author_url": "", "post_date": "12/11/2025 05:34:55", "content": "Hi, for data that has fewer than 7 body parts or doesn't match the 7 parts you selected, do you just discard it?
", "votes": null, "replies": [ { "id": 3371719, "author_name": "antoinemasq", "author_url": "", "post_date": "12/11/2025 16:06:31", "content": "No I keep everything and just replace those missing values by (0,0) and the model simply learns to ignore them. From my experiments, too much preprocessing will make the model perform worse.
", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ] } ] } ] } ] }, { "id": 3361990, "author_name": "yuanzhezhou", "author_url": "", "post_date": "12/04/2025 03:33:12", "content": "It is weird that nn underperform trees. Maybe all the top ranks are secretly using NN only?
", "votes": null, "replies": [ { "id": 3364431, "author_name": "woosungyoon", "author_url": "", "post_date": "12/06/2025 13:43:18", "content": "Your meta-awareness of Kaggle competitions is truly outstanding.
\nThe accuracy of your predictions seems to be extremely high.