Hi Guys,
\nI prepared some feature engineering notebooks for all of you to play around with!
\nThe Basic idea of those notebooks is to provide some simple \"cookbook\" for feature engineering functions you can all copy&paste from!
\nThey basically all have lists of feature engineering functions of different kinds and a demo for each function usage!
Neutralization
\nNeutralization is a technique used to battle overfitting. The main idea is to use a small linear model and subtract it's prediction from the target / features. Doing so, we \"neutralize\" the linear market trend effect on the features / target and allow our models to learn only from the \"unique\" information presented in the data.
Time Series Agg
\nA huge collection of Window-based aggregation functions (groupby->agg) that can be used via a simple interface to generate features.
Simple technical analysis feature extraction notebook that can be used as features for a model.
\nA bit more complex than the first notebook: This is a large collection of technical analysis indicators written in python (refactored by myself), all are tested and can be used as features.
\nFeatures that are based on common financial Volatility estimation methods.
\nHope it helps fellow Kagglers!
\nHappy Kaggling!
Hi Guys,
\nI prepared some feature engineering notebooks for all of you to play around with!
\nThe Basic idea of those notebooks is to provide some simple \"cookbook\" for feature engineering functions you can all copy&paste from!
\nThey basically all have lists of feature engineering functions of different kinds and a demo for each function usage!
Neutralization
\nNeutralization is a technique used to battle overfitting. The main idea is to use a small linear model and subtract it's prediction from the target / features. Doing so, we \"neutralize\" the linear market trend effect on the features / target and allow our models to learn only from the \"unique\" information presented in the data.
Time Series Agg
\nA huge collection of Window-based aggregation functions (groupby->agg) that can be used via a simple interface to generate features.
Simple technical analysis feature extraction notebook that can be used as features for a model.
\nA bit more complex than the first notebook: This is a large collection of technical analysis indicators written in python (refactored by myself), all are tested and can be used as features.
\nFeatures that are based on common financial Volatility estimation methods.
\nHope it helps fellow Kagglers!
\nHappy Kaggling!
Very nice collection of notebooks @yamqwe !
\nYou can also check out the new featurewiz library that creates most of these time series features and also helps you select the best features with just a couple of lines of code. It also runs on DASK which means that it can process millions of rows using parallelization. That could mean tremendous savings in time and effort.
\nLet me know after you have taken a look and may be you can write a notebook about it.
\nRam
@yamqwe :
\nHere is another notebook that might help you and others win Kaggle competitions since I have used this notebook extensively to score high ranks in the following competitions:
AnalyticsVidhya Hackathons: https://datahack.analyticsvidhya.com/contest/all/
\nMachine Hack Hackathons: https://www.machinehack.com/hackathon
\nQuick question, apart from derived features from High, Low, Volume, count none of it are given. So since train dataset is not updated and leadtime is 3 months in batches we will not able able to use any rolling time series features in the model as train data will have it filled as nulls.
", "rawMarkdown": "Quick question, apart from derived features from High, Low, Volume, count none of it are given. So since train dataset is not updated and leadtime is 3 months in batches we will not able able to use any rolling time series features in the model as train data will have it filled as nulls.", "votes": null }, { "id": "1629602", "postDate": "12/26/2021 10:51:11", "content": "Thank you 👍
", "rawMarkdown": "Thank you 👍", "votes": null }, { "id": "1630265", "postDate": "12/27/2021 05:51:42", "content": "can't they be accumulated during predictions?
", "rawMarkdown": "can't they be accumulated during predictions?", "votes": null } ], "comments": [ { "id": 1620057, "author_name": "rsesha", "author_url": "", "post_date": "12/16/2021 11:48:06", "content": "Very nice collection of notebooks @yamqwe !
\nYou can also check out the new featurewiz library that creates most of these time series features and also helps you select the best features with just a couple of lines of code. It also runs on DASK which means that it can process millions of rows using parallelization. That could mean tremendous savings in time and effort.
\nLet me know after you have taken a look and may be you can write a notebook about it.
\nRam
@yamqwe :
\nHere is another notebook that might help you and others win Kaggle competitions since I have used this notebook extensively to score high ranks in the following competitions:
AnalyticsVidhya Hackathons: https://datahack.analyticsvidhya.com/contest/all/
\nMachine Hack Hackathons: https://www.machinehack.com/hackathon
\nQuick question, apart from derived features from High, Low, Volume, count none of it are given. So since train dataset is not updated and leadtime is 3 months in batches we will not able able to use any rolling time series features in the model as train data will have it filled as nulls.
", "votes": null, "replies": [ { "id": 1630265, "author_name": "sapr3s", "author_url": "", "post_date": "12/27/2021 05:51:42", "content": "can't they be accumulated during predictions?
", "votes": null, "replies": [] } ] }, { "id": 1629602, "author_name": "aiizloli", "author_url": "", "post_date": "12/26/2021 10:51:11", "content": "Thank you 👍
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1616621": "Hi Guys, \n\nI prepared some feature engineering notebooks for all of you to play around with!\n\nThe Basic idea of those notebooks is to provide some simple \"cookbook\" for feature engineering functions you can all copy&paste from! \nThey basically all have lists of feature engineering functions of different kinds and a demo for each function usage! \n\n## The Notebooks:\n\n____\n\n\n**[Neutralization](https://www.kaggle.com/yamqwe/g-research-avoid-overfit-feature-neutralization/)**\nNeutralization is a technique used to battle overfitting. The main idea is to use a small linear model and subtract it's prediction from the target / features. Doing so, we \"neutralize\" the linear market trend effect on the features / target and allow our models to learn only from the \"unique\" information presented in the data.\n_____\n\n**[Time Series Agg](https://www.kaggle.com/yamqwe/features-all-time-series-aggregations-ever)**\nA huge collection of Window-based aggregation functions (groupby->agg) that can be used via a simple interface to generate features. \n\n_____\n**[Technical Analysis #1](https://www.kaggle.com/yamqwe/crypto-prediction-technical-analysis-features)**\n\nSimple technical analysis feature extraction notebook that can be used as features for a model.\n\n_____\n\n**[Technical Analysis #2](https://www.kaggle.com/yamqwe/crypto-prediction-technical-analysis-feats-2)**\n\nA bit more complex than the first notebook: This is a large collection of technical analysis indicators written in python (refactored by myself), all are tested and can be used as features. \n\n_____\n\n**[Volatility Features](https://www.kaggle.com/yamqwe/crypto-prediction-volatility-features)**\n\nFeatures that are based on common financial Volatility estimation methods. \n\n____\n\nHope it helps fellow Kagglers!\nHappy Kaggling!", "1620057": "Very nice collection of notebooks @yamqwe !\nYou can also check out the new [featurewiz library](https://github.com/AutoViML/featurewiz) that creates most of these time series features and also helps you select the best features with just a couple of lines of code. It also runs on DASK which means that it can process millions of rows using parallelization. That could mean tremendous savings in time and effort.\nLet me know after you have taken a look and may be you can write a notebook about it.\nRam", "1620071": "yamqwe :\nHere is another notebook that might help you and others win Kaggle competitions since I have used this notebook extensively to score high ranks in the following competitions:\n\n[Featurewiz Kaggle Template](https://www.kaggle.com/rsesha/featurewiz-kaggle-template)\n\n## The above Featurewiz+SimpleXGBModel Template has Received Top Scores in Multiple Hackathons - Current Ranks in the following Hackathons are:\nAnalyticsVidhya Hackathons: https://datahack.analyticsvidhya.com/contest/all/\n1. Big_Mart Sales Prediction Score: 1147 -- Rank 250 out of 41,361 = That's a Top <1% Rank!!\n1. Loan Status Predictions Score 0.791 -- Rank 850 out of 67,424 - Top 1.25% Rank\n\nMachine Hack Hackathons: https://www.machinehack.com/hackathon\n1. Machine Hack Flight Ticket Score 0.9389 -- Rank 165 out of 2723 - Top 6% Rank!\n1. Machine Hack Data Scientist Salary class Score 0.417 -- Rank 58 out of 1547 - Top 3.7% Rank! (Autoviml Score was 0.329 -- less than 0.417 of Featurewiz+Simple even though an NLP problem!)\n1. MCHACK Book Price NLP Score 0.7336 -- Rank 104 Autoviml NLP problem and should have done better", "1621953": "Quick question, apart from derived features from High, Low, Volume, count none of it are given. So since train dataset is not updated and leadtime is 3 months in batches we will not able able to use any rolling time series features in the model as train data will have it filled as nulls.", "1629602": "Thank you 👍", "1630265": "can't they be accumulated during predictions?" }, "source": "meta" }