Session-based Recommendation via Memory Network and Dwell-time Attention

Abstract

The session-based recommendation task predicts a user’s next interaction based on ongoing anonymous sessions. Previous studies mainly focused on improving click-through rates, while few studies have yielded further investigation on how to increase purchase conversion rates in session-based recommendations. However, the purchase rate improvement is a crucial metric to measure the performance of recommendation systems in e-commerce scenarios, which means that the prediction of users’ purchase behavior is of significant value. To this end, we build a novel model to predict what items people will be more likely to buy in the current session. The basic intuition of this paper is that within a session, the user’s dwell-time on each item reflects the interest level, and collaborative information from neighbor buy-sessions ((sessions that end with a purchased item) can be leveraged to boost the recommendation accuracy. We propose a new hybrid framework that leverages dwell-time information to model users’ actual interests. In addition, a memory network is employed to store and retrieve history buy-sessions as collaborative information. Finally, the representation of the current session and its neighbor sessions are combined to predict what items users are more likely to buy. Experimental results on two real e-commerce datasets show that the proposed model locates the actual demands (items to buy) more precisely than the state-of-art baselines.