Abstract
Session-Based Recommender Systems (SBRSs) process time-aware user–item interactions to capture users’ dynamic preferences. Most of the existing SBRSs mainly strive to improve recommendation accuracy by exploiting different deep learning techniques to represent each user on the basis of the most recent session. However, they generally ignore to capture the diversity preference at the session level. In this view, we consider both recommendation diversity and accuracy when generating recommendations. Particularly, we treat that users’ preferences towards diversity might be varied across users (sessions). Thus, we propose an end-to-end neural network model, namely Purpose Tendency-aware Diversified Strategy for Session-based Recommendation (PTDS-SR), where we design a Purpose Tendency Probability (PTP) module matching with a two-channel decoder to guide whether to recommend similar or diversified items given a session. We thus obtain a relatively personalized strategy for each session to exploit the diversity to facilitate recommendation accuracy regarding short-term user preferences (within a session). We compare our approach with representative, state-of-the-art baselines on three real-world datasets, in terms of recommendation accuracy, diversity and a comprehensive metric (considering both accuracy and diversity). Experimental results demonstrate the effectiveness of our approach over the state-of-the-art approaches. Meanwhile, with PTDS-SR, more purposeless sessions (e.g., searching products of various categories) will obtain more diverse recommendation list, and vice versa.
•An end to-end learning model, PTDS-SR, is proposed to improve recommendation accuracy and diversity for session-based recommendation.•PTDS-SR exploits the personalized diversity at the session level.•It designs a much more simplified definition to measure the diversity of a set for greedy algorithm.•Extensive experiments on real-world datasets have verified the effectiveness of our work.