A relation classification and aggregation algorithm for bipartite-type multi-relational heterogeneous graphs

Existing Heterogeneous Graph Neural Networks (HGNNs) are multi-oriented and single-relational heterogeneous graphs, and cannot effectively function on Bipartite-type Multi-relational Heterogeneous Graphs (BMHGs) with multiple relationships. At the same time, existing meta-path-based HGNNs cannot fully consider the differences between meta-paths during the aggregation process, and this difference is even more prominent in BMHGs. The main manifestation is that the number of neighbor nodes connected by various meta-relation paths differs significantly, causing some paths to carry too much noise information, which affects the algorithm performance. In order to solve the problem of the complex relationships in BMHG and the significant disparity in the number of neighbors between paths, this paper proposes a Relation Classification and Aggregation Algorithm for Bipartite-type Multi-Relational Heterogeneous Graphs (RCAA-BMHG). The RCAA-BMHG algorithm consists of three modules: the same-type aggregation module, the across-type aggregation module, and the cross-category feature aggregation layer, which perform differentiated processing of different types of association information between nodes in a Bipartite-type Multi-relational Heterogeneous Graph. Specifically, the same-type aggregation module first introduces a same-type node association filter to distinguish between the densely coupled path and the sparsely coupled path, and then uses a global average strategy and an adaptive weight allocation method to aggregate the information of the two types of coupled paths. The across-type aggregation module is filtered by an across-type node association filter, and then the weighted sum mechanism and the neighborhood feature propagation technology are used to aggregate the information of the two types of coupled paths. Finally, RCAA-BMHG uses a category-level attention mechanism to fuse the semantic and feature information of the same and cross types to generate the final node embedding for downstream tasks. Experimental verification shows that RCAA-BMHG not only performs feature aggregation and classification tasks when processing complex heterogeneous graph data, but also shows significant advantages over existing HGNNs algorithms on multiple evaluation metrics. The complete reproducible code and data have been published at: https://github.com/Dylanwsd24/RCAA-BMHG.