Global Interest Transfer Guided Session-based Recommendation

Abstract

Session-based recommendation aims to predict the next item based on the anonymous user’s clicked item sequence. Users’ interest in different content shifts regularly, and almost all of the current session based recommendation methods can’t capture the transfer relationship between interests, which can guide our prediction of the next item. This paper proposes an innovative method called Global Interest Transfer Guided Session based Recommendation(GITG), which uses global information to learn interest representations and transfer rules between interests to help the recommendation. In GITG, we parse sessions from two perspectives: (i)Interest: we learn the items’ interest representation by using the global neighbor set and learn the interests transfer relationship in the interest graph. (ii)Session: we learn the local embedding in the session graph and combine it with the global-post embedding. From these two perspectives, we can obtain interest representation and session representation, which provide high-value information for recommendation. Experiments show that GITG performs well on three real-world datasets.