Heterogeneous hypergraph representation learning for link prediction

Abstract

Heterogeneous graph representation learning gains popularity due to its powerful capabilities of feature extraction and numerous related algorithms have emerged for various downstream tasks in graph structural datasets. However, the interactions among nodes for the heterogeneous graphs in the real world often extend beyond individual pairs, excessive attention is payed on isolated pairwise connections. In this paper, we propose a novel framework of Heterogeneous Hypergraph Representation Learning method (HHRL) to capture high-order interactions for learning effective node representations of heterogeneous graphs. The method firstly organizes the heterogeneous connections as different hypergraphs. By modeling the heterogeneous connections, HHRL captures the rich structural and semantic information present in the graphs. Then, the graph neural network (GNN) is applied for each hypergraph to capture the interdependencies between nodes and their associated features. By utilizing GNN, HHRL can effectively learn expressive node representations that encode both the structural and feature information of the network. Finally, we concatenate the vectors from different hypergraphs to obtain the link representations. The experiments are conducted on five real dataset for link prediction and the results demonstrate the well performance of the proposed framework comparing to the existing baselines