Graph neural networks combined with multi-source graph information for session-based recommendation

Abstract:

Existing session-based recommendations with graph neural networks could capture the item's transition relationship by constructing graph structures from sessions. However, most graph neural networks and their improved models only consider the single transition relationship of items in the session when modeling sessions. As a result, a large amount of effective information is ignored, and the analysis of hidden correlations between items is lacking. Therefore, a session-based recommendation algorithm with graph neural network and multi-source graph information is proposed. In the algorithm, the users' repeat behavior information and item content related information are incorporated into the session graph modeling process, which effectively extracts the deeper complex transformation relationship of items, and aggregates it through linear transformations. In addition, an external attention mechanism is used to obtain the hidden association information of the session sequence items, making the generated session vectors more accurate. The experiments were performed on the real datasets: Yoochoose and Diginetica, and the results showed that the model outperformed the benchmark model. In particularly, it outperforms the state-of-the-art benchmark model GC-SAN, on average by 12.50% in terms of the MRR@20 evaluation metric, and can better predict user's next click items.