First of all, thank you for organizing a great competition. I would also like to thank everyone who participated in the competition with us and the four of us who worked together as a team.
\nSepcially thanks to my teammates: @trasibulo @virilo and @albert2017
We merge several types of models for gettint the final candidates. Each model contribute with different number of candidates and then we ensemble them over a ranking based weights to get the final 150 candidates. We tested 100 and 150, we noticed a slightly improvement with 150.
\nOur final top150 recall was:
\n clicks recall=0.7241
\n carts recall=0.5636
\n orders recall=0.7399
\n Overall Recall(top 150) = 0.6854
I started with pandas but finally I changed all my pipeline to polars.
\nI decided to do one model per item type: clicks, carts, orders
item_features
\nsession_features
\nuser_items_features
\nfeatures from candidates generation (probs, similarity, etc..)
We tried lgb and catboost;
\nAlthough lgb was slightly better on LB, catboost was faster so at the end I decided to go only with catboost.
I tried many things for improving validation score at the end but everything i tried didn't seem to help
\nUse carts predictions as feature for orders
\nStacking
\nDifferent models
\nTrain with all data
\nw2vec similarity features
Xgboost model and lgbm model. (comming soon)
\nBST Transformer.
\nhttps://www.kaggle.com/competitions/otto-recommender-system/discussion/384441
We tried both ensemble and stacking.
\nStacking: we didnt managed to make it work on LB (while we were getting a good validation score, it was not reflected on LB)
\nEnsemble: it seemed to be more stable but also didnt give us the expected boost.
\nAt the end our final submission was an ensemble of probabilities from 3 models: xgb,lgb,catboost
\nLB results-> single model (0.602 low); ensemble (0.602 high)
First of all, thank you for organizing a great competition. I would also like to thank everyone who participated in the competition with us and the four of us who worked together as a team.
\nSepcially thanks to my teammates: @trasibulo @virilo and @albert2017
We merge several types of models for gettint the final candidates. Each model contribute with different number of candidates and then we ensemble them over a ranking based weights to get the final 150 candidates. We tested 100 and 150, we noticed a slightly improvement with 150.
\nOur final top150 recall was:
\n clicks recall=0.7241
\n carts recall=0.5636
\n orders recall=0.7399
\n Overall Recall(top 150) = 0.6854
I started with pandas but finally I changed all my pipeline to polars.
\nI decided to do one model per item type: clicks, carts, orders
item_features
\nsession_features
\nuser_items_features
\nfeatures from candidates generation (probs, similarity, etc..)
We tried lgb and catboost;
\nAlthough lgb was slightly better on LB, catboost was faster so at the end I decided to go only with catboost.
I tried many things for improving validation score at the end but everything i tried didn't seem to help
\nUse carts predictions as feature for orders
\nStacking
\nDifferent models
\nTrain with all data
\nw2vec similarity features
Xgboost model and lgbm model. (comming soon)
\nBST Transformer.
\nhttps://www.kaggle.com/competitions/otto-recommender-system/discussion/384441
We tried both ensemble and stacking.
\nStacking: we didnt managed to make it work on LB (while we were getting a good validation score, it was not reflected on LB)
\nEnsemble: it seemed to be more stable but also didnt give us the expected boost.
\nAt the end our final submission was an ensemble of probabilities from 3 models: xgb,lgb,catboost
\nLB results-> single model (0.602 low); ensemble (0.602 high)
Polars is new to me, thanks for the mention @enric1296! I am sure this should be good as I can see some users using it frequently now-a-days.
\nCongrats for the great result! Best regards!
Congrats @enric1296
", "rawMarkdown": "Congrats @enric1296", "votes": null }, { "id": "2127745", "postDate": "02/03/2023 05:44:36", "content": "Well Done mate 👍
", "rawMarkdown": "Well Done mate 👍", "votes": null }, { "id": "2128137", "postDate": "02/03/2023 13:46:04", "content": "Congrats to all of you!
", "rawMarkdown": "Congrats to all of you!", "votes": null }, { "id": "2134442", "postDate": "02/08/2023 01:47:12", "content": "Congrats on the gold, well done.
", "rawMarkdown": "Congrats on the gold, well done.", "votes": null } ], "comments": [ { "id": 2127249, "author_name": "ravi20076", "author_url": "", "post_date": "02/02/2023 19:18:11", "content": "Polars is new to me, thanks for the mention @enric1296! I am sure this should be good as I can see some users using it frequently now-a-days.
\nCongrats for the great result! Best regards!
Congrats @enric1296
", "votes": null, "replies": [] }, { "id": 2127745, "author_name": "ayushnitb", "author_url": "", "post_date": "02/03/2023 05:44:36", "content": "Well Done mate 👍
", "votes": null, "replies": [] }, { "id": 2128137, "author_name": "indiratsirikhova", "author_url": "", "post_date": "02/03/2023 13:46:04", "content": "Congrats to all of you!
", "votes": null, "replies": [] }, { "id": 2134442, "author_name": "buumoo", "author_url": "", "post_date": "02/08/2023 01:47:12", "content": "Congrats on the gold, well done.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2127187": "First of all, thank you for organizing a great competition. I would also like to thank everyone who participated in the competition with us and the four of us who worked together as a team.\n\nSepcially thanks to my teammates: @trasibulo @virilo and @albert2017 \n\n## Candidates Selection\n\nWe merge several types of models for gettint the final candidates. Each model contribute with different number of candidates and then we ensemble them over a ranking based weights to get the final 150 candidates. We tested 100 and 150, we noticed a slightly improvement with 150.\n\n### Scores: \n\nOur final top150 recall was: