Disentangled hyperbolic representation learning for heterogeneous graphs

Qijie Bai, Changli Nie, Haiwei Zhang, Zhicheng Dou, Xiaojie Yuan

Abstract: Heterogeneous graphs have attracted considerable research interests in the past few years owing to their remarkable ability to represent complex real-world systems. However, the existing methods have two pain points in embedding heterogeneous graphs into low-dimensional spaces. (1) Heterogeneous graphs contain both structural and semantic information, yet these aspects are often mixed in original data, hence hindering effective representation learning. (2) Most real-world heterogeneous graphs exhibit power-law distributions over both the entire graphs and single edge types, and they suffer mismatch with Euclidean spaces, in which most existing methods are built. These two challenges require that representation methods should unmix the information while considering the distributions of entire graphs and single edge types. Herein, we propose Dis-H2GCN, a Disentangled Hyperbolic Heterogeneous Graph Convolutional Network. On the one hand, we leverage the mutual information minimization and discrimination maximization constraints to disentangle the semantic features from comprehensively learned representations by independent message propagation for each edge type, away from the pure structural features. On the other hand, the entire model is constructed upon hyperbolic geometry to narrow the gap between data distributions and representing spaces. We evaluate the performance of proposed Dis-H2GCN on five real-world heterogeneous graph datasets across two downstream tasks, node classification and link prediction, with the results demonstrating its superiority over state-of-the-art methods.