Adaptive Graph Clustering for Homophily using Generalized PageRank

Graph clustering is an important task for graph deep learning. Most graph clustering methods are developed on the basis of graph homophily assumption, which limited the performance of those methods over heterophilic graph data. In graph classification, many models like GPRGNN are proposed to work on both homophilic and heterophilic graph. However, these models cannot be easily transferred to the graph clustering scenario due to the lack of node labels. In the paper, we propose a novel universal clustering approach that combines the idea of general pagerank and contractive learning and can adaptively fit homophilic and heterophilic graph data. Extensive experiments on both homophilic and heterophilic benchmarks demonstrate our promising performance.