Upgrade to Pro — share decks privately, control downloads, hide ads and more …

H&M 23th place solution

H&M 23th place solution

kaggleのH&Mコンペの23位になりました。その解法の概略図です。

コンペURL: https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/
解法のdiscussion: https://www.kaggle.com/competitions/h-and-m-personalized-fashion-recommendations/discussion/324085

5473f177bf5562e68645e1a6e0428f9c?s=128

Kyohei Uto

May 27, 2022
Tweet

More Decks by Kyohei Uto

Other Decks in Science

Transcript

  1. H&M Personalized Fashion Recommendations Main process Most popular Most Popular

    items by segment(all, age, etc..) in the last 5 days ① Generate Item Candidates LGBM Ranker ② Create Features ⑤ Post Processing Copyright 2022 @kuto_bopro Candidates Strategies Features ③ Learn Ranking Model ④ Ensemble ・・・ ・・・ to cold start customers Use age based most popular items as predictions to all customers Remove the low offline sales ratio items from the customers who prefer offline. Blend 6 model predictions which is created by the different candidates and features Create below features by week, and merge to the generated candidates dataset. Dataset & CV Strategy train data week valid data week create candidates and features week 104w 103w 102w 101w 100w 0w Generate item candidates for each customers and each week with multiple strategies. transactions week User based CF Top similar items by user based CF and LightGCN Different color Different color items from past purchased Past purchased Past purchased items Item based CF Top similar items of past purchased items by item based CF Team: ZKMRD CF: Collaborative Filtering Best single model CV: 0.0390 (public) LB: 0.0325 Key points about dataset ・Use only candidate examples generated, not all positive examples ・Use customers with at least one positive example in the candidates ・Remove not for sales now items from candidates CF features ・score by item based CF ・score/rank by LightGCN (GCN based CF model) Article dynamic attributes ・trend value ・weekly popular ranking ・purchase count (1day, 2day ago, last week..) ・purchase count by segment (age, item group) ・popular rank by segment (age, item group) ・mean sales channel Article static attributes ・basic attributes in articles.csv ・Bert sentence vector Customer dynamic attributes ・purchase count ・last purchase flag ・Days/Week since last purchase ・price/discount of purchased items ・mean sales channel ・purchase rate by item segment ・repurchase rate Customer static attributes ・basic attributes in customers.csv Private LB: 0.0329 (23th)