Robust Graph Collaborative Filtering Algorithm Based on Hierarchical Attention

Abstract

Learning the embedding representation of users and items is the core of the collaborative filtering algorithm. In recent years, the graph neural network (GNN) has been applied to the recommendation field due to its excellent performance. However, in the process of GNN iteratively aggregating neighbor information, the occasional noise in the graph structure will transmit errors to neighbor nodes along with the aggregation process, which will worsen the embedding representation of other nodes. Noise information is ubiquitous in real life. Therefore, while mining high-order collaborative information in depth, improving the robustness of the GNN model is also an important factor that needs to be considered in the recommendation task. Based on the above problems, this paper proposes a robust graph collaborative filtering algorithm based on hierarchical attention, which includes node-level and graph-level attention. Node-level attention performs preference aggregation and occasional noise information filtering on different neighbor nodes by learning the attention coefficients of different neighbor nodes; graph-level attention performs fusion and secondary filtering of occasional noise information on different deep graph embeddings by learning the attention coefficients of different dimensional nodes. The node-level and graph-level attention can fully realize the noise reduction of the graph structure in the deep propagation process. While ensuring that the nodes encode high-order collaborative information, it minimizes the noise information carried. Extensive experimental results on three data sets show that the recommendation algorithm is better than the existing mainstream recommendation algorithm in all evaluation indicators.