Hybrid-HGT: A unified deep learning framework for community detection in heterogeneous attributed graphs

Abstract

Identifying communities in heterogeneous attributed graphs remains a significant challenge for networks consisting of multi-membership nodes and multi-relations, attribute-rich information, and very few labels. Here, we present a combined HGT-based framework incorporating contrastive representation learning and deep clustering with multi-round training via pseudo-labels. This enables us to jointly utilize both structural connections and semantic similarities in order to yield more informative node embeddings and discover interesting community structures. Experiments on DBLP and IMDB benchmark datasets demonstrate that the proposed framework is able to achieve competitive performance, verifying its effectiveness for community detection in real-world heterogeneous networks. In addition, the iterative pseudo-labeling mechanism forms a powerful learning strategy that progressively increases supervision by conferring confident predictions on currently unlabeled nodes, while the contrastive and clustering objectives facilitate representation discrimination and community cohesion, enabling our model to better deal with label scarcity, heterogeneous dependencies, and overlapping semantics in practical, complex, attributed networks.