I built on this excellent baseline and made the following changes:
\nhttps://www.kaggle.com/code/ravaghi/social-action-recognition-in-mice-xgboost
\nThank you @ravaghi.
\nExcluded non-test datasets from training: removed CalMS21, MABe22, and CRIM13 since they do not appear in the test set. (This made CV correlate better with LB.)
Trained models per lab: trained XGBoost and CatBoost separately for each lab. (LightGBM performed worse for some reason.)
Removed AdaptableSnail 25 fps from training.
Bidirectional EMA smoothing: applied EMA in both forward and backward directions, and searched for the best smoothing timescale.
Ego-centric features: used body_center as the origin (for labs without body_center, I used the mean of the annotation coordinates), rotated coordinates so the motion direction aligns with the positive Y-axis, and generated ego-centric features.
Feature pruning: removed features whose importance was zero in both CatBoost and XGBoost.
Fair tie-breaking when multiple classes exceed thresholds: argmax() is not fair under class imbalance, so I used a Z-score–based margin:
z = (prob - threshold) / sigma, where sigma is the standard deviation of OOF predictions.Ensembling: averaged EMA-smoothed probabilities across models and used the average threshold for decisions; when multiple classes competed, selected the one with the highest mean Z-score across models.
max_gap and removing spikes via min_duration.rear, submit, approach).I built on this excellent baseline and made the following changes:
\nhttps://www.kaggle.com/code/ravaghi/social-action-recognition-in-mice-xgboost
\nThank you @ravaghi.
\nExcluded non-test datasets from training: removed CalMS21, MABe22, and CRIM13 since they do not appear in the test set. (This made CV correlate better with LB.)
Trained models per lab: trained XGBoost and CatBoost separately for each lab. (LightGBM performed worse for some reason.)
Removed AdaptableSnail 25 fps from training.
Bidirectional EMA smoothing: applied EMA in both forward and backward directions, and searched for the best smoothing timescale.
Ego-centric features: used body_center as the origin (for labs without body_center, I used the mean of the annotation coordinates), rotated coordinates so the motion direction aligns with the positive Y-axis, and generated ego-centric features.
Feature pruning: removed features whose importance was zero in both CatBoost and XGBoost.
Fair tie-breaking when multiple classes exceed thresholds: argmax() is not fair under class imbalance, so I used a Z-score–based margin:
z = (prob - threshold) / sigma, where sigma is the standard deviation of OOF predictions.Ensembling: averaged EMA-smoothed probabilities across models and used the average threshold for decisions; when multiple classes competed, selected the one with the highest mean Z-score across models.
max_gap and removing spikes via min_duration.rear, submit, approach).\"Congrats @shu01! I'm really curious, how much of a boost did you get from the ego-centric features?
\nAlso, I managed to improve my score by nearly 0.02 using max_gap and min_duration for post-processing, so I'm wondering if you could share the details of how you handled this part?\"
", "rawMarkdown": "\"Congrats @shu01! I'm really curious, how much of a boost did you get from the ego-centric features?\n\nAlso, I managed to improve my score by nearly 0.02 using max_gap and min_duration for post-processing, so I'm wondering if you could share the details of how you handled this part?\"", "votes": null }, { "id": "3377423", "postDate": "12/16/2025 09:38:55", "content": "The egocentric features gave about a +0.005 gain on both CV and LB. (I double-checked they were implemented correctly using egocentric coordinates.)
\nAs for post-processing: after converting frame-wise predictions into segments, I used OOF predictions to grid-search action-specific values of max_gap and min_duration (in seconds). However, applying this post-processing did not change the LB score at all. I suspect it was largely overfitting to CV. I originally introduced it because I noticed the sample submission had more rows than I expected given the video durations, but since I was already applying EMA smoothing, this extra step might have been unnecessary (or even too aggressive).
", "rawMarkdown": "The egocentric features gave about a +0.005 gain on both CV and LB. (I double-checked they were implemented correctly using egocentric coordinates.)\n\nAs for post-processing: after converting frame-wise predictions into segments, I used OOF predictions to grid-search action-specific values of max_gap and min_duration (in seconds). However, applying this post-processing did not change the LB score at all. I suspect it was largely overfitting to CV. I originally introduced it because I noticed the sample submission had more rows than I expected given the video durations, but since I was already applying EMA smoothing, this extra step might have been unnecessary (or even too aggressive).", "votes": null }, { "id": "3377436", "postDate": "12/16/2025 10:09:16", "content": "\"Thanks for the answer. The relative coordinate system is indeed an impressive feature extraction method.
\nPerhaps the key difference lies in how these parameters were derived: my min_gap and min_duration were based on statistics from the ground truth labels of the entire training set, and I used frames as the unit.\"
", "rawMarkdown": "\"Thanks for the answer. The relative coordinate system is indeed an impressive feature extraction method.\n\nPerhaps the key difference lies in how these parameters were derived: my min_gap and min_duration were based on statistics from the ground truth labels of the entire training set, and I used frames as the unit.\"", "votes": null } ], "comments": [ { "id": 3377409, "author_name": "ybzjdbc", "author_url": "", "post_date": "12/16/2025 09:11:16", "content": "\"Congrats @shu01! I'm really curious, how much of a boost did you get from the ego-centric features?
\nAlso, I managed to improve my score by nearly 0.02 using max_gap and min_duration for post-processing, so I'm wondering if you could share the details of how you handled this part?\"
", "votes": null, "replies": [ { "id": 3377423, "author_name": "shu01jp", "author_url": "", "post_date": "12/16/2025 09:38:55", "content": "The egocentric features gave about a +0.005 gain on both CV and LB. (I double-checked they were implemented correctly using egocentric coordinates.)
\nAs for post-processing: after converting frame-wise predictions into segments, I used OOF predictions to grid-search action-specific values of max_gap and min_duration (in seconds). However, applying this post-processing did not change the LB score at all. I suspect it was largely overfitting to CV. I originally introduced it because I noticed the sample submission had more rows than I expected given the video durations, but since I was already applying EMA smoothing, this extra step might have been unnecessary (or even too aggressive).
", "votes": null, "replies": [ { "id": 3377436, "author_name": "ybzjdbc", "author_url": "", "post_date": "12/16/2025 10:09:16", "content": "\"Thanks for the answer. The relative coordinate system is indeed an impressive feature extraction method.
\nPerhaps the key difference lies in how these parameters were derived: my min_gap and min_duration were based on statistics from the ground truth labels of the entire training set, and I used frames as the unit.\"
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3377383": "I built on this excellent baseline and made the following changes:\n\nhttps://www.kaggle.com/code/ravaghi/social-action-recognition-in-mice-xgboost\n\nThank you @ravaghi.\n\n\n* **Excluded non-test datasets from training**: removed **CalMS21**, **MABe22**, and **CRIM13** since they do not appear in the test set. (This made CV correlate better with LB.)\n* **Trained models per lab**: trained **XGBoost** and **CatBoost** separately for each lab. (LightGBM performed worse for some reason.)\n* **Removed AdaptableSnail 25 fps from training**.\n* **Bidirectional EMA smoothing**: applied EMA in both forward and backward directions, and searched for the best smoothing timescale.\n* **Ego-centric features**: used **body_center** as the origin (for labs without body_center, I used the mean of the annotation coordinates), rotated coordinates so the motion direction aligns with the **positive Y-axis**, and generated ego-centric features.\n* **Feature pruning**: removed features whose importance was **zero in both CatBoost and XGBoost**.\n* **Fair tie-breaking when multiple classes exceed thresholds**: `argmax()` is not fair under class imbalance, so I used a Z-score–based margin:\n\n * `z = (prob - threshold) / sigma`, where **sigma is the standard deviation of OOF predictions**.\n * Selected the class with the highest z-score.\n* **Ensembling**: averaged **EMA-smoothed probabilities** across models and used the **average threshold** for decisions; when multiple classes competed, selected the one with the highest **mean Z-score across models**.\n\n### What didn’t help (for me)\n\n* Post-processing after inference: filling with `max_gap` and removing spikes via `min_duration`.\n* Left-right flip augmentation for GBDT (**CV↑ but LB↓**).\n* Handling AdaptableSnail 25 fps based on discussion ideas (scaling frames by **1.2×**, removing `rear`, `submit`, `approach`).", "3377409": "\"Congrats @shu01! I'm really curious, how much of a boost did you get from the ego-centric features?\n\nAlso, I managed to improve my score by nearly 0.02 using max_gap and min_duration for post-processing, so I'm wondering if you could share the details of how you handled this part?\"", "3377423": "The egocentric features gave about a +0.005 gain on both CV and LB. (I double-checked they were implemented correctly using egocentric coordinates.)\n\nAs for post-processing: after converting frame-wise predictions into segments, I used OOF predictions to grid-search action-specific values of max_gap and min_duration (in seconds). However, applying this post-processing did not change the LB score at all. I suspect it was largely overfitting to CV. I originally introduced it because I noticed the sample submission had more rows than I expected given the video durations, but since I was already applying EMA smoothing, this extra step might have been unnecessary (or even too aggressive).", "3377436": "\"Thanks for the answer. The relative coordinate system is indeed an impressive feature extraction method.\n\nPerhaps the key difference lies in how these parameters were derived: my min_gap and min_duration were based on statistics from the ground truth labels of the entire training set, and I used frames as the unit.\"" }, "source": "meta" }