SNGN:Similar Neighbors Graph Transition Network for Session-based Recommendation

The target of session-based recommendation is to predict the next item to be clicked by the user based on anonymous sequences. The current mainstream session recommendation model is to build the session into a graph to capture item transitions within it. However, this only gives information about the current session, ignoring complex project information in other sessions. If other sessions are blindly added to the model, noise may be introduced, so we propose a new method to calculate adjacent sessions. In this paper, we propose a novel method, namely a similar neighbors Graph Transition Network for Session-based Recommendation (SNGN), to model item transitions within not only the current session but also the neighbor sessions. Specifically, first, we get the neighbor session graph through our new neighbor session computing method. Then input the current session graph and neighbor session graph into the perceptron to get the corresponding representation. Finally, the two representations are fused to obtain the user preference representation. Experiments on real-world datasets demonstrate that SNGN outperforms other state-of-the-art methods.