Okay almost 1 month in the competition and I would like to share some of the common strategies that are working for me . Hope this will be helpful to the new participants in this competition and also improve models of existing participants. That being said please feel free to share what strategies and tricks are working for you in the comments its the kaggle way :) .
\n1.Cross Validation:
\nGroupKfold based on subject offers stable cv , the CV and LB are well correlated to each other and consistent . Hence sticking to models that are consistent with CV and LB is the most crucial aspect while building stable models single models and even ensemble of models
2. Feature Engineering :
\nPhysics based and kinematics features like acceleration , rotation , angular velocity , derivatives offers most boost to the score as compared to the other features a good collection of features can be found in @rktqwe kernel here . However adding too many features having multicollinearity can lead to problem and can decrease score in models.
3. Models:
\nNeural Networks like 1D Cnn and 2 Branch neural networks are dominating this task the single models are itself too strong . Didnt found a good performing kernel using Gradient Boosting models . However it would be interesting to see other models apart from 1D Cnn and branch neural networks that can beat 0.77 -0.78 single model score with good cv . It would be interesting to ensemble such models
4.IMU , TOH and THM Signals:
\nAs said in competition data description there are half of the hidden-test sequences are recorded with IMU only for such sequences training and predicting using model trained on IMU data is much more beneficial than using models trained on TOH and THM signals .
5. Demographics information and features :
\nI tried using demographics information and features but the model score doesn't improves and in some cases it deteriorates model scores may be demographics information is just noise for this task .
6. Postprocessing :
\nTraining model using Label Smoothing works well for me however down the line it would be interesting to use some other post processing techniques while making final predictions
7. Ensembling :
\nAs of now havent tried ensemble models as best performing models are architecturally similar 1D CNN or branch CNN . In future we can try ensembling models like different variants of NN or GBDT models we can ensemble them with branch NN models to improve the results.
This are some of the insights and findings from my side if you have any and would like to share please let me know in the comments below .
\nThanks ,
\nHappy Kaggling :) .
Okay almost 1 month in the competition and I would like to share some of the common strategies that are working for me . Hope this will be helpful to the new participants in this competition and also improve models of existing participants. That being said please feel free to share what strategies and tricks are working for you in the comments its the kaggle way :) .
\n1.Cross Validation:
\nGroupKfold based on subject offers stable cv , the CV and LB are well correlated to each other and consistent . Hence sticking to models that are consistent with CV and LB is the most crucial aspect while building stable models single models and even ensemble of models
2. Feature Engineering :
\nPhysics based and kinematics features like acceleration , rotation , angular velocity , derivatives offers most boost to the score as compared to the other features a good collection of features can be found in @rktqwe kernel here . However adding too many features having multicollinearity can lead to problem and can decrease score in models.
3. Models:
\nNeural Networks like 1D Cnn and 2 Branch neural networks are dominating this task the single models are itself too strong . Didnt found a good performing kernel using Gradient Boosting models . However it would be interesting to see other models apart from 1D Cnn and branch neural networks that can beat 0.77 -0.78 single model score with good cv . It would be interesting to ensemble such models
4.IMU , TOH and THM Signals:
\nAs said in competition data description there are half of the hidden-test sequences are recorded with IMU only for such sequences training and predicting using model trained on IMU data is much more beneficial than using models trained on TOH and THM signals .
5. Demographics information and features :
\nI tried using demographics information and features but the model score doesn't improves and in some cases it deteriorates model scores may be demographics information is just noise for this task .
6. Postprocessing :
\nTraining model using Label Smoothing works well for me however down the line it would be interesting to use some other post processing techniques while making final predictions
7. Ensembling :
\nAs of now havent tried ensemble models as best performing models are architecturally similar 1D CNN or branch CNN . In future we can try ensembling models like different variants of NN or GBDT models we can ensemble them with branch NN models to improve the results.
This are some of the insights and findings from my side if you have any and would like to share please let me know in the comments below .
\nThanks ,
\nHappy Kaggling :) .
thanks for sharing this!
", "rawMarkdown": "thanks for sharing this!", "votes": null }, { "id": "3253160", "postDate": "07/24/2025 05:50:17", "content": "Thank you!
", "rawMarkdown": "Thank you!", "votes": null } ], "comments": [ { "id": 3237812, "author_name": "vaishnavimudaliar", "author_url": "", "post_date": "07/01/2025 10:23:04", "content": "thanks for sharing this!
", "votes": null, "replies": [] }, { "id": 3253160, "author_name": "jyagaimo292", "author_url": "", "post_date": "07/24/2025 05:50:17", "content": "Thank you!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3236592": "Okay almost 1 month in the competition and I would like to share some of the common strategies that are working for me . Hope this will be helpful to the new participants in this competition and also improve models of existing participants. **That being said please feel free to share what strategies and tricks are working for you in the comments its the kaggle way :) .**\n\n**1.Cross Validation**:\nGroupKfold based on subject offers stable cv , the CV and LB are well correlated to each other and consistent . Hence sticking to models that are consistent with CV and LB is the most crucial aspect while building stable models single models and even ensemble of models\n\n**2. Feature Engineering** :\nPhysics based and kinematics features like acceleration , rotation , angular velocity , derivatives offers most boost to the score as compared to the other features a good collection of features can be found in @rktqwe kernel [here ](https://www.kaggle.com/code/rktqwe/lb-0-77-linear-accel-tf-bilstm-gru-attention). **However adding too many features having multicollinearity can lead to problem and can decrease score in models.**\n\n**3. Models**:\nNeural Networks like 1D Cnn and 2 Branch neural networks are dominating this task the single models are itself too strong . Didnt found a good performing kernel using Gradient Boosting models . However it would be interesting to see other models apart from 1D Cnn and branch neural networks that can beat 0.77 -0.78 single model score with good cv . It would be interesting to ensemble such models\n\n**4.IMU , TOH and THM Signals**:\nAs said in competition data description there are half of the hidden-test sequences are recorded with IMU only for such sequences training and predicting using model trained on IMU data is much more beneficial than using models trained on TOH and THM signals .\n\n**5. Demographics information and features** : \nI tried using demographics information and features but the model score doesn't improves and in some cases it deteriorates model scores may be demographics information is just noise for this task .\n\n**6. Postprocessing** : \nTraining model using Label Smoothing works well for me however down the line it would be interesting to use some other post processing techniques while making final predictions\n\n**7. Ensembling** : \nAs of now havent tried ensemble models as best performing models are architecturally similar 1D CNN or branch CNN . In future we can try ensembling models like different variants of NN or GBDT models we can ensemble them with branch NN models to improve the results.\n\nThis are some of the insights and findings from my side if you have any and would like to share please let me know in the comments below .\n\nThanks ,\nHappy Kaggling :) .", "3237812": "thanks for sharing this!", "3253160": "Thank you!" }, "source": "meta" }