PEPNet: Parameter and Embedding Personalized Network for Infusing with Personalized Prior Information

• DOI: 10.48550/arxiv.2302.01115
• Publication Date: 2023-01-01
• Authors: Jianxin Chang; Chenbin Zhang; Yiqun Hui; Dewei Leng; Yanan Niu; Yang Song; Kun Gai
• Keywords: FOS: Computer and information sciences; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; Information Retrieval (cs.IR); Computer Science - Information Retrieval

With the increase of content pages and interactive buttons in online services such as online-shopping video-watching websites, industrial-scale recommender systems face challenges multi-domain multi-task recommendations. The core recommendation is to accurately capture user interests multiple scenarios given behaviors. In this paper, we propose a plug-and-play \textit{\textbf{P}arameter \textbf{E}mbedding \textbf{P}ersonalized \textbf{Net}work (\textbf{PEPNet})} for recommendation. PEPNet takes personalized prior information input dynamically scales bottom-level Embedding top-level DNN hidden units through gate mechanisms. \textit{Embedding Personalized Network (EPNet)} performs selection on fuse features with different importance users domains. \textit{Parameter (PPNet)} executes modification parameters balance targets sparsity tasks. We have made series special engineering optimizations combining Kuaishou training framework deployment environment. By infusing parameters, tailored each individual obtains significant performance gains, improvements exceeding 1\% task metrics across deployed apps, serving over 300 million every day.