Learning sequential and general interests via a joint neural model for session-based recommendation

Abstract

Session-based recommendation is to predict the next action of a user based on his current behavior session. Recent studies have focused on designing recurrent neural networks to capture representation of users sequential behaviors from a large number of anonymous sessions. Although these neural models have greatly improved recommendation performance compared with traditional approaches, they have overemphasized the sequential information of sessions, that is, the order of items in each browsing session. We argue that besides the order of items, the set of items in each session can also reflect the general interests of current session and can also be exploited to augment recommendations. To this end, this paper proposes a joint neural model, called SGINM, for jointly learning sequential and general interests of each session for session-based recommendation. In SGINM, we design an attentive recurrent neural network which not only captures self-adaptive weighed representation of each subsequence in a session but also learns a global representation for the whole session. Furthermore, the SGINM adopts a multi-layer neural network with residual connections to learn the general interests of each session. The SGINM uses a softmax layer to jointly decode the two types of interests and output a ranking vector of recommended items for each new session. The proposed model has been extensively experimented with the benchmark datasets from the RecSys Challenge 2015 and CIKM Cup 2016. Compared with the state-of-the-art models, the proposed SGINM achieves better recommendation performance in terms of higher predicative accuracy and mean reciprocal rank in almost all experiment cases.