Enhanced Multi-Head Self-Attention Graph Neural Networks for Session-based Recommendation

Session-based recommendation aims to predict following user behaviors based on short-term anonymous sessions. Although there are many types of existing recommendation models, it is still a challenging problem to dig out the profound relationship between users and items from short sessions. Therefore, complex transformation patterns and the dynamic evolution of user interests are considered in the session, and this paper proposes a new method, named Enhanced Multi-Head Self-Attention Graph Neural Networks for Session-based Recommendation (EMSAGNN). The model first converts sequence data into graph structure and feeds them into graph neural networks to dynamically learn the complex transition patterns and capture user preferences. Then, an enhanced multi-head self-attention network further learns sessions to capture rich hidden information in items. Finally, with the learned information, EMSAGNN calculates the probability scores of different items to recommend more suitable items for users. We conduct extensive experiments and comparisons on two public e-commerce datasets. The experimental results show that our proposed model is superior to the state-of-the-art methods.