Global Feature Extraction Graph Neural Networks for Session Recommendation

The recommender system based on session sequence is mainly used for the recommendation task when the user information is scarce. Usually, the user session is expressed as an ordered sequence of items. However, learning user preferences from ordered sequences is a complex problem in deep learning based recommender system. This paper proposes a novel model, the Global Feature Extraction Graph Neural Network for Session Recommendation (GFE-GNN), which transforms the session sequence from sequential representation to graphical representation. While obtaining the sequence order, the global graph feature extraction is used to obtain the potential sequence of items, so as to learn the complex preferences of users. Finally, the problem of predicting the next item of the session is transformed into a graph classification problem through the aggregation layer. Experimental results on real datasets show that GFE-GNN is superior to the most advanced session recommendation methods.