Congratulations to all the winners 🥇 and thanks to Kaggle for a wonderful competition. Compiling the gold medal solutions from the competition below with short summary of the approaches.
\nWill be updated as we get more solutions from the winners.
", "messages": [ { "id": "1783113", "postDate": "05/10/2022 06:15:26", "content": "Congratulations to all the winners 🥇 and thanks to Kaggle for a wonderful competition. Compiling the gold medal solutions from the competition below with short summary of the approaches.
\nWill be updated as we get more solutions from the winners.
", "rawMarkdown": "Congratulations to all the winners 🥇 and thanks to Kaggle for a wonderful competition. Compiling the gold medal solutions from the competition below with short summary of the approaches.\n\n* [1st place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324070) - Used a two step approach - candidate generation followed by ranking model. Candidate generation played a vital role and spent good time on improving the recall of top candidates. Created a wide range of features for ranking model and used multiple cumulative features as many people do not have transactions in last months. Ensembled 5 LightGBM and 7 CatBoost models to get the final score.\n* [2nd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324197) - Used the two step approach as well - candidate generation followed by ranking. Used attributed of users, items, embeddings, co-occurrences and random graph walk over iterm-user graph to generate candidates. Wrote the feature engineering part in Rust for faster execution. Used lambdarankmap objective for LGBMRanker and it performed better than the lambdarank (which uses NDCG objective).\n* [3rd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324129) - Used a similar two stage approach. Couple of notable things are - a) Rank of the candidate by the given candidate generation strategy as a feature input and b) User to item similarity obtained from BPR matrix factorization.\n* [4th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324094) - Followed the candidate generation model to improve the recall and then ranking model approach. 5 fold CV with 3 folds as training data. Developed 4 candidate generation models and used LightGBM with lambda rank and DCN model as ranking models. Used text and image features for cold-start problems.\n* [5th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324098) - Used 21 different methods to create the candidate generation. Created article embeddings using different data and methods. Calibrated different recall methods as they have different positive ratio. Used a single lightgbm ranker model as final model.\n* [6th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324075) - Also used the candidate generation followed by ranking approach. Top features are count, gap, discount, tfidf and item similarity ones. Used last 9 weeks for training with the last week as validation. Best single model is a catboost model and final model is an ensemble of several models. [part-2](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324278)\n* [8th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324185) - Used the two step approach. Used 7 fold CV by using last 7 weeks as validation sample. 3 weeks before the validation is used as test for LGBMRanker models.\n* [9th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324127) - Used last week as validation. Trained LightGBM model with around 300 features. Generated 400 candidates per user in the first stage. Ensembled 4 models to get the final score.\n* [10th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324223) - Very high CV-LB correlation even at higher scores. Used the two step approach similar to other top teams. Used a simple sales prediction model to detect products that are phasing in or out, which is one of the most important feature.\n* [11th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324084) - Also followed the candidate generation (multiple ways) and ranking strategy. CatBoost with Yetirank is the top model. Used 6, 8 and 12 weeks for training and last week as validation. Final model is an ensemble of 4 models trained with 4 different weeks and different params. \n* [12th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324310)\n* [13th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324207)\n* [16th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324158) - Splitted the data into three groups based on number of transactions made by the customer and took different approaches for different customers. Used lightgbm ranker and binary classification models.\n\nWill be updated as we get more solutions from the winners.", "votes": null }, { "id": "1783114", "postDate": "05/10/2022 06:15:57", "content": "Other top solutions shared from the competition are
\nVery nice compilation. Thanks for doing this!
", "rawMarkdown": "Very nice compilation. Thanks for doing this!", "votes": null } ], "comments": [ { "id": 1783114, "author_name": "sudalairajkumar", "author_url": "", "post_date": "05/10/2022 06:15:57", "content": "Other top solutions shared from the competition are
\nVery nice compilation. Thanks for doing this!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1783113": "Congratulations to all the winners 🥇 and thanks to Kaggle for a wonderful competition. Compiling the gold medal solutions from the competition below with short summary of the approaches.\n\n* [1st place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324070) - Used a two step approach - candidate generation followed by ranking model. Candidate generation played a vital role and spent good time on improving the recall of top candidates. Created a wide range of features for ranking model and used multiple cumulative features as many people do not have transactions in last months. Ensembled 5 LightGBM and 7 CatBoost models to get the final score.\n* [2nd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324197) - Used the two step approach as well - candidate generation followed by ranking. Used attributed of users, items, embeddings, co-occurrences and random graph walk over iterm-user graph to generate candidates. Wrote the feature engineering part in Rust for faster execution. Used lambdarankmap objective for LGBMRanker and it performed better than the lambdarank (which uses NDCG objective).\n* [3rd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324129) - Used a similar two stage approach. Couple of notable things are - a) Rank of the candidate by the given candidate generation strategy as a feature input and b) User to item similarity obtained from BPR matrix factorization.\n* [4th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324094) - Followed the candidate generation model to improve the recall and then ranking model approach. 5 fold CV with 3 folds as training data. Developed 4 candidate generation models and used LightGBM with lambda rank and DCN model as ranking models. Used text and image features for cold-start problems.\n* [5th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324098) - Used 21 different methods to create the candidate generation. Created article embeddings using different data and methods. Calibrated different recall methods as they have different positive ratio. Used a single lightgbm ranker model as final model.\n* [6th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324075) - Also used the candidate generation followed by ranking approach. Top features are count, gap, discount, tfidf and item similarity ones. Used last 9 weeks for training with the last week as validation. Best single model is a catboost model and final model is an ensemble of several models. [part-2](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324278)\n* [8th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324185) - Used the two step approach. Used 7 fold CV by using last 7 weeks as validation sample. 3 weeks before the validation is used as test for LGBMRanker models.\n* [9th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324127) - Used last week as validation. Trained LightGBM model with around 300 features. Generated 400 candidates per user in the first stage. Ensembled 4 models to get the final score.\n* [10th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324223) - Very high CV-LB correlation even at higher scores. Used the two step approach similar to other top teams. Used a simple sales prediction model to detect products that are phasing in or out, which is one of the most important feature.\n* [11th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324084) - Also followed the candidate generation (multiple ways) and ranking strategy. CatBoost with Yetirank is the top model. Used 6, 8 and 12 weeks for training and last week as validation. Final model is an ensemble of 4 models trained with 4 different weeks and different params. \n* [12th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324310)\n* [13th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324207)\n* [16th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324158) - Splitted the data into three groups based on number of transactions made by the customer and took different approaches for different customers. Used lightgbm ranker and binary classification models.\n\nWill be updated as we get more solutions from the winners.", "1783114": "Other top solutions shared from the competition are\n\n* [17th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324595)\n* [22nd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324224), [part-2](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324152)\n* [23rd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324085)\n* [25th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324293)\n* [26th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324350)\n* [30th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324411)\n* [46th place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324205)\n* [53rd place solution](https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324076)", "1783895": "Very nice compilation. Thanks for doing this!" }, "source": "meta" }