頂級會議論壇PPT：岑宇闊.pdf

1、1 Controllable Multi-Interest Framework for Recommendation Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, Jie Tang Tsinghua KEG & Alibaba DAMO Academy Paper: https:/arxiv.org/abs/2005.09347 Github: https:/ 2 Background Rapid development of e-commerce Personalized recommender systems 3 R

2、ecommender System Industrial recommender systems: matching stage ranking stage The matching stage: retrieve top-N candidate items The ranking stage: sort the candidate items by scores Matching: candidate generation Ranking hundreds Item Corpus item 1 item 2 item n 4 Introduction to two Stages Matchi

3、ng stage: f(user behaviors) = user embedding dot(user embedding, item embedding) need fast k-nearest neighbor search models: YouTube DNN, GRU4REC Ranking stage: f(user behaviors, item embedding) = 0/1 models: Wide&Deep, DeepFM 5 K-Nearest Neighbor Search Fast k-nearest neighbor search method is needed for the matching stage, . Faiss: a library for efficient similarity search of dense vectors Faiss