Multi-behavior Graph Neural Networks for Session-based Recommendation

Abstract

Session-based recommendation aims to predict user behaviors based on short-term anonymous sessions. Given the singleness of the training data of the existing session recommendation model, it is not easy to learn from a simple session to a complex transition pattern. To obtain complete session information and complex transitions of items, we propose a novel method, named Multi-behavior Graph Neural Networks for Session-based Recommendation (MBGNN). Specifically, this study uses multiple behavior sequences to construct a session graph, and input the session graph into the graph neural network to learn the complex transition patterns between items to obtain the global information about the session. Then, according to the different importance of the various sessions in the recommendation process, feature fusion is carried out to obtain a more detailed local representation. Finally, the learned information is input into the prediction layer to calculate the probability score of each item to make more accurate recommendations for users. Extensive experiments and comparisons were conducted on two real datasets. The experimental results show that our proposed model is superior to the state-of-the-art methods.