Feature Engineering is Critical and Moreover an optimize way of generating features is also important !
Loop Feature Engineering Helps alot to come up with features fast enough and avoid merges ! Hence good ideas is to get comfortable with loops
Correctness of user and content is an important feature , think about ways you can come up with correctness features other than those provided in public kernels(Note:Use Loops and Dictionaries Here ).
Rolling and Lags Features will be important , but Our Dataframe needs to be sorted to come up with them ! Moreover it is better to write implementation in Numpy rather than using inbuilt pandas functions ! One Simple Hint is given here
Modelling LightGBM works like Magic ! One should Focus on Getting most out of LightGBM and then move onto NN models ! As seen in M5 Forecasting Accuracy , Competitions The 1st Solution , and the 4th Solution , LightGBM fused with robust validation and Feature Engineering can get very high ranks.
Rather than using Strong Hyperparameter Tuning and spending alot of time on it , it will be better to do it once or twice ! If we do lot of hyperparameter Tuning it may so happen that Data may overfit on validation set and may not generalize on External Test set unknown to us .
Lot of Ensembling and Merges should be avoided to overcome the submission timeout error !
Feature Engineering is Critical and Moreover an optimize way of generating features is also important !
Loop Feature Engineering Helps alot to come up with features fast enough and avoid merges ! Hence good ideas is to get comfortable with loops
Correctness of user and content is an important feature , think about ways you can come up with correctness features other than those provided in public kernels(Note:Use Loops and Dictionaries Here ).
Rolling and Lags Features will be important , but Our Dataframe needs to be sorted to come up with them ! Moreover it is better to write implementation in Numpy rather than using inbuilt pandas functions ! One Simple Hint is given here
Modelling LightGBM works like Magic ! One should Focus on Getting most out of LightGBM and then move onto NN models ! As seen in M5 Forecasting Accuracy , Competitions The 1st Solution , and the 4th Solution , LightGBM fused with robust validation and Feature Engineering can get very high ranks.
Rather than using Strong Hyperparameter Tuning and spending alot of time on it , it will be better to do it once or twice ! If we do lot of hyperparameter Tuning it may so happen that Data may overfit on validation set and may not generalize on External Test set unknown to us .
Lot of Ensembling and Merges should be avoided to overcome the submission timeout error !
M5 Forecasting Accuracy,aha
", "rawMarkdown": "M5 Forecasting Accuracy,aha" }, { "id": 1115798, "postDate": "2020-12-16T15:27:09.600Z", "content": "thanks for sharing ! Correctness is Important！
", "rawMarkdown": "thanks for sharing ! Correctness is Important！" }, { "id": 1110942, "postDate": "2020-12-13T08:17:42.330Z", "content": "great insights @sayedathar11
", "rawMarkdown": "great insights @sayedathar11 " }, { "id": 1121823, "postDate": "2020-12-21T23:59:14.010Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1111219, "postDate": "2020-12-13T14:35:26.080Z", "content": "thank you for sharing
", "rawMarkdown": "thank you for sharing" }, { "id": 1111023, "postDate": "2020-12-13T10:03:47.103Z", "content": "thanks for sharing !
", "rawMarkdown": "thanks for sharing !" }, { "id": 1109680, "postDate": "2020-12-12T00:26:15.660Z", "content": "thanks for sharing !
", "rawMarkdown": "thanks for sharing !" } ], "comments": [ { "id": 1118411, "author_name": "zyliang1994", "author_url": "", "post_date": "2020-12-19T03:43:26.393000", "content": "M5 Forecasting Accuracy,aha
", "votes": 0, "replies": [] }, { "id": 1115798, "author_name": "Jingzhe_Zhu", "author_url": "", "post_date": "2020-12-16T15:27:09.600000", "content": "thanks for sharing ! Correctness is Important！
", "votes": 0, "replies": [] }, { "id": 1110942, "author_name": "Ritesh Kumar", "author_url": "", "post_date": "2020-12-13T08:17:42.330000", "content": "great insights @sayedathar11
", "votes": 0, "replies": [] }, { "id": 1121823, "author_name": "", "author_url": "", "post_date": "2020-12-21T23:59:14.010000", "content": "", "votes": 0, "replies": [] }, { "id": 1111219, "author_name": "Sharlto Cope", "author_url": "", "post_date": "2020-12-13T14:35:26.080000", "content": "thank you for sharing
", "votes": 0, "replies": [] }, { "id": 1111023, "author_name": "Paul Wang", "author_url": "", "post_date": "2020-12-13T10:03:47.103000", "content": "thanks for sharing !
", "votes": 0, "replies": [] }, { "id": 1109680, "author_name": "Mnka", "author_url": "", "post_date": "2020-12-12T00:26:15.660000", "content": "thanks for sharing !
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1109404": "1. Feature Engineering is Critical and Moreover an optimize way of generating features is also important !\n\n2. Loop Feature Engineering Helps alot to come up with features fast enough and avoid merges ! Hence good ideas is to get comfortable with loops\n\n3. **Correctness** of user and content is an important feature , think about ways you can come up with correctness features other than those provided in public kernels(**Note:Use Loops and Dictionaries Here** ).\n\n4. Rolling and Lags Features will be important , but Our Dataframe needs to be sorted to come up with them ! Moreover it is better to write implementation in Numpy rather than using inbuilt pandas functions ! One Simple Hint is given [here](https://www.kaggle.com/c/riiid-test-answer-prediction/discussion/202714)\n\n5. Modelling LightGBM works like Magic ! One should Focus on Getting most out of LightGBM and then move onto NN models ! As seen in **M5 Forecasting Accuracy** , Competitions The [1st Solution](https://www.kaggle.com/c/m5-forecasting-accuracy/discussion/163684) , and the [4th Solution](https://www.kaggle.com/c/m5-forecasting-accuracy/discussion/163216) , LightGBM fused with robust validation and Feature Engineering can get very high ranks.\n\n6. Rather than using Strong Hyperparameter Tuning and spending alot of time on it , it will be better to do it once or twice ! If we do lot of hyperparameter Tuning it may so happen that Data may overfit on validation set and may not generalize on External Test set unknown to us .\n\n7. Lot of Ensembling and Merges should be avoided to overcome the submission timeout error !\n\n", "1118411": "M5 Forecasting Accuracy,aha", "1115798": "thanks for sharing ! Correctness is Important！", "1110942": "great insights @sayedathar11 ", "1121823": "", "1111219": "thank you for sharing", "1111023": "thanks for sharing !", "1109680": "thanks for sharing !" } }