Improved Global Context Graph Neural Networks for Session-based Recommendation

Session-based recommendation is designed to make recommendations for anonymous or non-logged users based on insession data. The existing research work usually only models a single item in a conversation as the smallest unit, ignoring the representation of the item. This research aimed to optimize the use of context information across the global spectrum by analyzing the interaction patterns of items within a session, along with any potential correlation between the various conversational adjacencies. The proposed model would learn from both the session graph and the global graph, and then combine this data to obtain an embedded representation of the items. The attention mechanism would be applied to attain the embedded representation of the session in order to predict future interaction behaviors. To compare with current session-based recommendation algorithms, the model in this paper is recommended to perform better than the previous model.