Dual-View Fusion of Heterogeneous Information Network Embedding for Recommendation

Abstract

Heterogeneous Information Networks (HINs) contain rich semantic information due to their involvement of multiple types of nodes and edges. Heterogeneous network embedding is used to analyze HINs by embedding network information in low-dimensional node representations. However, existing heterogeneous embedding methods either ignore the implicit topological relationships between distant nodes or neglect nodes features and meta-paths information disparities, which reflects that extracting HIN embeddings from a single view may lead to incomplete information extraction. In order to make the information extraction more complete, we propose a dual-view fusion heterogeneous information network embedding method (DFHE) for recommendation tasks. Specifically, it extracts effective features from HINs from both the remote topology view and the semantic aggregation view: the remote topology view uses a meta-graph-guided random walk to capture the topological relationships between remote nodes and learns embeddings through a graph convolutional network (GCN) encoder, while the semantic aggregation view uses an attention mechanism to learn the importance of different meta-paths, node relationships, and aggregates the semantic information of each meta-path. Experimental results on two real-world network datasets demonstrate an enhancement in recommendation task performance under the application of DFHE, compared to the baseline. This improvement persists even when some meta-paths are deleted, thereby verifying the methods effectiveness.