Self-supervised global graph neural networks with enhance-attention for session-based recommendation

Abstract

Session-based recommendation is a challenging task which predicts the next click based on the short-term behavior of anonymous users. Compared to other recommendation models, session-based recommendations are more difficult due to the limited amount of available data, which is also data sparsity. To solve the problem, we induce self-supervised learning, which can be incorporated into network training by constructing real samples from raw data. It generates self-supervised signals and maximizes the mutual information of session expressions learned. In addition, we propose an enhanced attention module called Enhance-attention. It combines knowledge from global-level graphs and session-level graphs and enhances the intent representation of sessions using Transformer. In this paper, we propose a new approach, called EAT-SGNN, that is able to predict the next click in a more granular way using all items in the session. The model is augmented by self-supervised learning that generates supervised signals. The model is evaluated on three public datasets: Tmall, Nowplaying, and Diginetica. According to the experimental results, the proposed method achieves state-of-the-art performance. All the code and datasets are publicly available on https://github.com/ch30git798/EAT-SGNN.git.