{ "id": 408012, "title": "Similar compition and solutions share👀", "url": "/competitions/tlvmc-parkinsons-freezing-gait-prediction/discussion/408012", "author_name": "", "post_date": "2023-05-09T04:01:07.501756600Z", "votes": 22, "comment_count": 3, "views": 0, "content": "

hi fellow kagglers:
\nWhen I browse the website, I find a similar competition: University of Liverpool - Ion Switching, in which also need us to build a multiclass predictor or regressor on several high frequency time series data.
\nHere are some of the solutions I've compiled. In general terms, most top solutions use wavenet, very few use tree-based model and HMM to build there model. And nearly all of then pay lot attention on data processing as the particularity of data(noise, data drift, etc)

\n

ml solution

\n

Preprocessing and 26 Model GBDT Ensemble, this solution only use tree-based model, but the key is not just feature engineering,

\n
    \n
  1. before he build models, he conducted extensive data analysis on the raw data, carried out corresponding raw data processing based on the analysis, and created models for different trends of data. The author also provided the corresponding ideas and code, which you can click to view.
    2. \n
  2. As for feature engineering, it was relatively simple. He used less than twenty features. As a contrast, this notebook Ion Switching - AdvFE, LGB, Wavenet, ConfMatrix organized most of the feature engineering during the compition and submitted, but the improvement in results was not significant.
    3. \n
\n

However, we can also refer to the feature engineering approach used in this competition :

\n
    \n
  1. shift_pastorlead, this [kernal](Ion Switching Advanced EDA and Prediction) includ 1~3
    2. \n
  2. rollingwindow_pastorlead,
    3. \n
  3. groupbykeycol_stats,
    4. \n
  4. target_encoding(use bin to x(in that compition is \"signal\"; in this mybe \"AccV\", \"AccML\", \"AccAP\") and use category-encoders later)
    5. \n
  5. calc_gradients, calc_low_pass, calc_ewm described in this kernel
    \nin this compition, the most useful feature is shift_pastorlead because of the way the raw data is generated, if you are interested in this, you can check this [notebook](On Markov Chains and the Competition Data)
    6. \n
\n

NN solution

\n

the most used baseline model: Wavenet
\n11th Place Solution,the author presented his entire training process and provided the code for some key steps, including keyfeature he used, model structure, optimization for metrics.
\n2nd place solution shared his code. He didn't say much, but we can still learn from the code he provided.

", "messages": [ { "id": "2251078", "postDate": "05/09/2023 04:01:07", "content": "

hi fellow kagglers:
\nWhen I browse the website, I find a similar competition: University of Liverpool - Ion Switching, in which also need us to build a multiclass predictor or regressor on several high frequency time series data.
\nHere are some of the solutions I've compiled. In general terms, most top solutions use wavenet, very few use tree-based model and HMM to build there model. And nearly all of then pay lot attention on data processing as the particularity of data(noise, data drift, etc)

\n

ml solution

\n

Preprocessing and 26 Model GBDT Ensemble, this solution only use tree-based model, but the key is not just feature engineering,

\n
    \n
  1. before he build models, he conducted extensive data analysis on the raw data, carried out corresponding raw data processing based on the analysis, and created models for different trends of data. The author also provided the corresponding ideas and code, which you can click to view.
    2. \n
  2. As for feature engineering, it was relatively simple. He used less than twenty features. As a contrast, this notebook Ion Switching - AdvFE, LGB, Wavenet, ConfMatrix organized most of the feature engineering during the compition and submitted, but the improvement in results was not significant.
    3. \n
\n

However, we can also refer to the feature engineering approach used in this competition :

\n
    \n
  1. shift_pastorlead, this [kernal](Ion Switching Advanced EDA and Prediction) includ 1~3
    2. \n
  2. rollingwindow_pastorlead,
    3. \n
  3. groupbykeycol_stats,
    4. \n
  4. target_encoding(use bin to x(in that compition is \"signal\"; in this mybe \"AccV\", \"AccML\", \"AccAP\") and use category-encoders later)
    5. \n
  5. calc_gradients, calc_low_pass, calc_ewm described in this kernel
    \nin this compition, the most useful feature is shift_pastorlead because of the way the raw data is generated, if you are interested in this, you can check this [notebook](On Markov Chains and the Competition Data)
    6. \n
\n

NN solution

\n

the most used baseline model: Wavenet
\n11th Place Solution,the author presented his entire training process and provided the code for some key steps, including keyfeature he used, model structure, optimization for metrics.
\n2nd place solution shared his code. He didn't say much, but we can still learn from the code he provided.

", "rawMarkdown": "hi fellow kagglers:\n\nWhen I browse the website, I find a similar competition: [University of Liverpool - Ion Switching](https://www.kaggle.com/competitions/liverpool-ion-switching/overview), in which also need us to build a multiclass predictor or regressor on several high frequency time series data.\n\nHere are some of the solutions I've compiled. In general terms, most top solutions use wavenet, very few use tree-based model and HMM to build there model. And nearly all of then pay lot attention on data processing as the particularity of data(noise, data drift, etc)\n\n## ml solution\n\n[Preprocessing and 26 Model GBDT Ensemble](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/153750), this solution only use tree-based model, but the key is not just feature engineering, \n1. before he build models, he conducted extensive data analysis on the raw data, carried out corresponding raw data processing based on the analysis, and created models for different trends of data. The author also provided the corresponding ideas and code, which you can click to view. \n2. As for feature engineering, it was relatively simple. He used less than twenty features. As a contrast, this notebook [Ion Switching - AdvFE, LGB, Wavenet, ConfMatrix](https://www.kaggle.com/code/vbmokin/ion-switching-advfe-lgb-wavenet-confmatrix?scriptVersionId=40518221) organized most of the feature engineering during the compition and submitted, but the improvement in results was not significant.\n \nHowever, we can also refer to the feature engineering approach used in this competition : \n1. shift_pastorlead, this [kernal]([Ion Switching Advanced EDA and Prediction](https://www.kaggle.com/code/gpreda/ion-switching-advanced-eda-and-prediction/notebook)) includ 1~3\n2. rollingwindow_pastorlead, \n3. groupbykeycol_stats, \n4. target_encoding(use bin to x(in that compition is \"signal\"; in this mybe \"AccV\", \"AccML\", \"AccAP\") and use [category-encoders](https://pypi.org/project/category-encoders/) later)\n5. calc_gradients, calc_low_pass, calc_ewm described in this [kernel](https://www.kaggle.com/code/ragnar123/single-model-lgbm/notebook)\n\nin this compition, the most useful feature is shift_pastorlead because of the way the raw data is generated, if you are interested in this, you can check this [notebook]([On Markov Chains and the Competition Data](https://www.kaggle.com/code/friedchips/on-markov-chains-and-the-competition-data/notebook))\n\n## NN solution\n\nthe most used baseline model: [Wavenet](https://www.kaggle.com/siavrez/wavenet-keras)\n\n[11th Place Solution](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/154264),the author presented his entire training process and provided the code for some key steps, including keyfeature he used, model structure, optimization for metrics.\n\n[2nd place solution](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/153991) shared his code. He didn't say much, but we can still learn from the code he provided.", "votes": null }, { "id": "2251399", "postDate": "05/09/2023 09:50:24", "content": "

Great effort!

", "rawMarkdown": "Great effort!", "votes": null }, { "id": "2254301", "postDate": "05/10/2023 20:09:16", "content": "

Great job !

", "rawMarkdown": "Great job !", "votes": null }, { "id": "2254480", "postDate": "05/11/2023 03:28:29", "content": "

Additional insight I want share about the baseline wavenet:

\n
    \n
  1. writer use features and raw x_data to the model. the features were proved useful in lgbm model
    2. \n
  2. he use data augments tech by using \"np.flip\"
    3. \n
  3. 17th Private & 17th Public Place Solution was most similar to to the baseline and also he shared his full kernel😄
    4. \n
", "rawMarkdown": "Additional insight I want share about the baseline wavenet:\n1. writer use features and raw x_data to the model. the features were proved useful in lgbm model\n2. he use data augments tech by using \"np.flip\"\n3. [17th Private & 17th Public Place Solution](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/153829) was most similar to to the baseline and also he shared his full kernel😄", "votes": null } ], "comments": [ { "id": 2251399, "author_name": "exjustice", "author_url": "", "post_date": "05/09/2023 09:50:24", "content": "

Great effort!

", "votes": null, "replies": [] }, { "id": 2254301, "author_name": "gianpaolobulleddu", "author_url": "", "post_date": "05/10/2023 20:09:16", "content": "

Great job !

", "votes": null, "replies": [] }, { "id": 2254480, "author_name": "roger92", "author_url": "", "post_date": "05/11/2023 03:28:29", "content": "

Additional insight I want share about the baseline wavenet:

\n
    \n
  1. writer use features and raw x_data to the model. the features were proved useful in lgbm model
    2. \n
  2. he use data augments tech by using \"np.flip\"
    3. \n
  3. 17th Private & 17th Public Place Solution was most similar to to the baseline and also he shared his full kernel😄
    4. \n
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2251078": "hi fellow kagglers:\n\nWhen I browse the website, I find a similar competition: [University of Liverpool - Ion Switching](https://www.kaggle.com/competitions/liverpool-ion-switching/overview), in which also need us to build a multiclass predictor or regressor on several high frequency time series data.\n\nHere are some of the solutions I've compiled. In general terms, most top solutions use wavenet, very few use tree-based model and HMM to build there model. And nearly all of then pay lot attention on data processing as the particularity of data(noise, data drift, etc)\n\n## ml solution\n\n[Preprocessing and 26 Model GBDT Ensemble](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/153750), this solution only use tree-based model, but the key is not just feature engineering, \n1. before he build models, he conducted extensive data analysis on the raw data, carried out corresponding raw data processing based on the analysis, and created models for different trends of data. The author also provided the corresponding ideas and code, which you can click to view. \n2. As for feature engineering, it was relatively simple. He used less than twenty features. As a contrast, this notebook [Ion Switching - AdvFE, LGB, Wavenet, ConfMatrix](https://www.kaggle.com/code/vbmokin/ion-switching-advfe-lgb-wavenet-confmatrix?scriptVersionId=40518221) organized most of the feature engineering during the compition and submitted, but the improvement in results was not significant.\n \nHowever, we can also refer to the feature engineering approach used in this competition : \n1. shift_pastorlead, this [kernal]([Ion Switching Advanced EDA and Prediction](https://www.kaggle.com/code/gpreda/ion-switching-advanced-eda-and-prediction/notebook)) includ 1~3\n2. rollingwindow_pastorlead, \n3. groupbykeycol_stats, \n4. target_encoding(use bin to x(in that compition is \"signal\"; in this mybe \"AccV\", \"AccML\", \"AccAP\") and use [category-encoders](https://pypi.org/project/category-encoders/) later)\n5. calc_gradients, calc_low_pass, calc_ewm described in this [kernel](https://www.kaggle.com/code/ragnar123/single-model-lgbm/notebook)\n\nin this compition, the most useful feature is shift_pastorlead because of the way the raw data is generated, if you are interested in this, you can check this [notebook]([On Markov Chains and the Competition Data](https://www.kaggle.com/code/friedchips/on-markov-chains-and-the-competition-data/notebook))\n\n## NN solution\n\nthe most used baseline model: [Wavenet](https://www.kaggle.com/siavrez/wavenet-keras)\n\n[11th Place Solution](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/154264),the author presented his entire training process and provided the code for some key steps, including keyfeature he used, model structure, optimization for metrics.\n\n[2nd place solution](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/153991) shared his code. He didn't say much, but we can still learn from the code he provided.", "2251399": "Great effort!", "2254301": "Great job !", "2254480": "Additional insight I want share about the baseline wavenet:\n1. writer use features and raw x_data to the model. the features were proved useful in lgbm model\n2. he use data augments tech by using \"np.flip\"\n3. [17th Private & 17th Public Place Solution](https://www.kaggle.com/competitions/liverpool-ion-switching/discussion/153829) was most similar to to the baseline and also he shared his full kernel😄" }, "source": "meta" }