Session-based Recommendation with Preference Interaction

Abstract

Graph neural network have achieved great success in session-based recommendation. Recently, some works have achieved improvement by incorporating income and outcome adjacent matrices to generate global and local preferences, and directly model the two preferences to build session representation. However, firstly, we observe that the income matrix and outcome matrix of a session have no strong relevance, and their concatenation may introduce noise for building two preferences. Secondly, we find the global and local preferences can benefit from each other, and collaborative information from neighborhood sessions may help to improve recommendation performance. Therefore, we propose a session-based recommendation with preference interaction from separate income adjacent matrix and outcome adjacent matrix framework, which includes two parallel modules: An Income Session Representation Encoder (ISE) and an Outcome Session Representation En-coder (OSE). A fusion gating mechanism is introduced to balance the importance of session representations resulting from ISE and OSE. The experimental results show that our model obviously outperforms other state-of-the-art methods on Yoochoose and Diginetica datasets.