MIGC: Multi-intent Graph Contrastive Learning in Recommendation

Abstract

Contrastive learning has been highly successful with computer vision and natural language processing. It can effectively address the under-sample situation. Contrastive learning has also been successfully implemented in recommender systems. It can not only address the problem of the small number of samples but also improve the learning impact of long-tailed data. Recommender systems contain large amounts of graph data. Graph neural networks are good at learning graph node representations. Through the neighbor information in the graph, it is possible to understand the potential intention of the user. Contrastive learning mainly includes sequence-based and graph-based contrastive learning in recommender systems. Currently, the modeling of both sequence contrastive learning and graph comparison learning in recommender systems is based on the user's single intent. However, the user's behavior consists of multiple intents. This paper proposes a new method which is named MIGC for modeling of user's numerous intents. Graph contrastive learning is introduced into the recommendation system recall algorithm and User's multi-interest modeling. This approach not only learns multiple users' intents but also improves the representation of long-tail data. Firstly, we construct a bipartite graph from user-to-item behavior data. Secondly, the multi-intents of users are a model of the graph. Finally, vector representations of users and items are obtained through contrastive learning of graph neural networks for vector recall in recommender systems. The experiments in this paper used the public dataset MovieLens and the private dataset e-commerce. And both offline and online have achieved a certain improvement. This study aims to start a new approach to users' multi-intent recall.