Many thanks to the organizers—this challenge offered a rich playground of modeling ideas and yielded plenty of hard-won lessons
\nTL;DR
\nI used subject-wise GroupStratifiedKFold CV (5 folds for analysis; 20-fold bagging for the final submission), careful handedness normalization (mirror left→right), and 180° correction for two reversed-worn subjects. The modeling stack combined WaveNet, ResNet, U-Net, and SqueezeFormer on 96-step sequences, trained with sample weights, time-series augmentations, and EMA weight smoothing. Because some sequences lacked rotation (quat) features, I trained four model variants (IMU-only vs IMU+THM+ToF; with/without accel-only). At inference, I gated examples to the appropriate variant when rotation features were missing. This gating yielded a +0.0015 improvement on the Private LB, which kept me in the gold zone.
Pre-processing & normalization
\nCross-validation
\nCV scheme: GroupStratifiedKFold by subject to preserve class proportions while preventing subject leakage.
Folds: I ran experiments with 5 folds (reporting CV), then built the final ensemble via 20-fold bagging (diverse splits/seeds under the same subject-grouped principle).
Why this matters: subjects have distinct motion signatures; mixing a subject across train/valid inflates CV. Grouped CV produced stable CV↔LB alignment and reduced overfitting risk.
Features
\nI combined competitive public feature ideas with additional temporal/physical descriptors:
Public baseline features (adopted and pruned from community notebooks)
Temporal gradients
Absolute values
Quaternion-derived features
Model family (multi-architecture ensemble)
\nAll models consumed 96-step sequences with identical train/valid splits.
Four training variants (to handle missing rotations)
\nInference routing: when rotation/quat-derived features were missing, I used (2) or (4) accordingly. Otherwise (1) or (3). This simple gate materially improved final reliability on mixed-modality test data.
\nTraining details
\nMany thanks to the competition organizers and the Kaggle community for resources and discussions
", "messages": [ { "id": "3284515", "postDate": "09/08/2025 10:45:33", "content": "Many thanks to the organizers—this challenge offered a rich playground of modeling ideas and yielded plenty of hard-won lessons
\nTL;DR
\nI used subject-wise GroupStratifiedKFold CV (5 folds for analysis; 20-fold bagging for the final submission), careful handedness normalization (mirror left→right), and 180° correction for two reversed-worn subjects. The modeling stack combined WaveNet, ResNet, U-Net, and SqueezeFormer on 96-step sequences, trained with sample weights, time-series augmentations, and EMA weight smoothing. Because some sequences lacked rotation (quat) features, I trained four model variants (IMU-only vs IMU+THM+ToF; with/without accel-only). At inference, I gated examples to the appropriate variant when rotation features were missing. This gating yielded a +0.0015 improvement on the Private LB, which kept me in the gold zone.
Pre-processing & normalization
\nCross-validation
\nCV scheme: GroupStratifiedKFold by subject to preserve class proportions while preventing subject leakage.
Folds: I ran experiments with 5 folds (reporting CV), then built the final ensemble via 20-fold bagging (diverse splits/seeds under the same subject-grouped principle).
Why this matters: subjects have distinct motion signatures; mixing a subject across train/valid inflates CV. Grouped CV produced stable CV↔LB alignment and reduced overfitting risk.
Features
\nI combined competitive public feature ideas with additional temporal/physical descriptors:
Public baseline features (adopted and pruned from community notebooks)
Temporal gradients
Absolute values
Quaternion-derived features
Model family (multi-architecture ensemble)
\nAll models consumed 96-step sequences with identical train/valid splits.
Four training variants (to handle missing rotations)
\nInference routing: when rotation/quat-derived features were missing, I used (2) or (4) accordingly. Otherwise (1) or (3). This simple gate materially improved final reliability on mixed-modality test data.
\nTraining details
\nMany thanks to the competition organizers and the Kaggle community for resources and discussions
", "rawMarkdown": "Many thanks to the organizers—this challenge offered a rich playground of modeling ideas and yielded plenty of hard-won lessons\n\n**TL;DR**\nI used subject-wise GroupStratifiedKFold CV (5 folds for analysis; 20-fold bagging for the final submission), careful handedness normalization (mirror left→right), and 180° correction for two reversed-worn subjects. The modeling stack combined WaveNet, ResNet, U-Net, and SqueezeFormer on 96-step sequences, trained with sample weights, time-series augmentations, and EMA weight smoothing. Because some sequences lacked rotation (quat) features, I trained four model variants (IMU-only vs IMU+THM+ToF; with/without accel-only). At inference, I gated examples to the appropriate variant when rotation features were missing. This gating yielded a +0.0015 improvement on the Private LB, which kept me in the gold zone.\n\n**Pre-processing & normalization**\n- **Handedness normalization**: mirrored **left-hand** recordings to **right-hand ** coordinates to unify kinematics.\n- **Reversed-wear correction**: two subjects wore the device reversed; I applied a **180°** rotation to bring their frames in line.\n- **Sequence windowing**: fixed **sequence length = 96** samples per example.\n- **Missing rotation features** : some sequences lacked quaternions/derived rotation features (Na). This motivated the model variants below and the inference router.\n\n\n**Cross-validation**\n- **CV scheme:** GroupStratifiedKFold by subject to preserve class proportions while preventing subject leakage.\n- **Folds:** I ran experiments with 5 folds (reporting CV), then built the final ensemble via 20-fold bagging (diverse splits/seeds under the same subject-grouped principle).\n- **Why this matters:** subjects have distinct motion signatures; mixing a subject across train/valid inflates CV. Grouped CV produced stable CV↔LB alignment and reduced overfitting risk.\n\n- \n**Features**\nI combined competitive public feature ideas with additional temporal/physical descriptors:\n- Public baseline features (adopted and pruned from community notebooks)\n- Temporal gradients\n- Absolute values\n- Quaternion-derived features\n\n\n**Model family (multi-architecture ensemble)**\nAll models consumed **96-step sequences** with identical train/valid splits.\n- WaveNet\n- ResNet\n- U-Net\n- SqueezeFormer\n\n\nFour training variants (to handle missing rotations)\n1. IMU-only (all IMU features).\n2. IMU-only (accelerometer features only).\n3. IMU + THM + ToF (all IMU features).\n4. IMU(acc-only) + THM + ToF.\n\n**Inference routing:** when rotation/quat-derived features were missing, I used **(2) or (4)** accordingly. Otherwise **(1) or (3).** This simple gate materially improved final reliability on mixed-modality test data.\n\n\n**Training details**\n- **Loss & weighting:** class imbalance addressed via sample weights\n- **Augmentation (time series):** light jitter, time-shift, and scale perturbations; mirror/rotation ops conformed to physically plausible transforms.\n- **EMA weight smoothing:** maintained an Exponential Moving Average of model parameters; EMA weights used for validation/inference to stabilize loss.\n\n\nMany thanks to the competition organizers and the Kaggle community for resources and discussions", "votes": null }, { "id": "3286073", "postDate": "09/09/2025 07:52:48", "content": "Congrats on solo gold and well deserved grandmaster rank =)
", "rawMarkdown": "Congrats on solo gold and well deserved grandmaster rank =)", "votes": null }, { "id": "3286081", "postDate": "09/09/2025 08:03:39", "content": "Thank you @datafan07 :)
", "rawMarkdown": "Thank you @datafan07 :)", "votes": null } ], "comments": [ { "id": 3286073, "author_name": "datafan07", "author_url": "", "post_date": "09/09/2025 07:52:48", "content": "Congrats on solo gold and well deserved grandmaster rank =)
", "votes": null, "replies": [ { "id": 3286081, "author_name": "davutpolat", "author_url": "", "post_date": "09/09/2025 08:03:39", "content": "Thank you @datafan07 :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3284515": "Many thanks to the organizers—this challenge offered a rich playground of modeling ideas and yielded plenty of hard-won lessons\n\n**TL;DR**\nI used subject-wise GroupStratifiedKFold CV (5 folds for analysis; 20-fold bagging for the final submission), careful handedness normalization (mirror left→right), and 180° correction for two reversed-worn subjects. The modeling stack combined WaveNet, ResNet, U-Net, and SqueezeFormer on 96-step sequences, trained with sample weights, time-series augmentations, and EMA weight smoothing. Because some sequences lacked rotation (quat) features, I trained four model variants (IMU-only vs IMU+THM+ToF; with/without accel-only). At inference, I gated examples to the appropriate variant when rotation features were missing. This gating yielded a +0.0015 improvement on the Private LB, which kept me in the gold zone.\n\n**Pre-processing & normalization**\n- **Handedness normalization**: mirrored **left-hand** recordings to **right-hand ** coordinates to unify kinematics.\n- **Reversed-wear correction**: two subjects wore the device reversed; I applied a **180°** rotation to bring their frames in line.\n- **Sequence windowing**: fixed **sequence length = 96** samples per example.\n- **Missing rotation features** : some sequences lacked quaternions/derived rotation features (Na). This motivated the model variants below and the inference router.\n\n\n**Cross-validation**\n- **CV scheme:** GroupStratifiedKFold by subject to preserve class proportions while preventing subject leakage.\n- **Folds:** I ran experiments with 5 folds (reporting CV), then built the final ensemble via 20-fold bagging (diverse splits/seeds under the same subject-grouped principle).\n- **Why this matters:** subjects have distinct motion signatures; mixing a subject across train/valid inflates CV. Grouped CV produced stable CV↔LB alignment and reduced overfitting risk.\n\n- \n**Features**\nI combined competitive public feature ideas with additional temporal/physical descriptors:\n- Public baseline features (adopted and pruned from community notebooks)\n- Temporal gradients\n- Absolute values\n- Quaternion-derived features\n\n\n**Model family (multi-architecture ensemble)**\nAll models consumed **96-step sequences** with identical train/valid splits.\n- WaveNet\n- ResNet\n- U-Net\n- SqueezeFormer\n\n\nFour training variants (to handle missing rotations)\n1. IMU-only (all IMU features).\n2. IMU-only (accelerometer features only).\n3. IMU + THM + ToF (all IMU features).\n4. IMU(acc-only) + THM + ToF.\n\n**Inference routing:** when rotation/quat-derived features were missing, I used **(2) or (4)** accordingly. Otherwise **(1) or (3).** This simple gate materially improved final reliability on mixed-modality test data.\n\n\n**Training details**\n- **Loss & weighting:** class imbalance addressed via sample weights\n- **Augmentation (time series):** light jitter, time-shift, and scale perturbations; mirror/rotation ops conformed to physically plausible transforms.\n- **EMA weight smoothing:** maintained an Exponential Moving Average of model parameters; EMA weights used for validation/inference to stabilize loss.\n\n\nMany thanks to the competition organizers and the Kaggle community for resources and discussions", "3286073": "Congrats on solo gold and well deserved grandmaster rank =)", "3286081": "Thank you @datafan07 :)" }, "source": "meta" }