PageRank Attributed Graph Subspace Clustering
Abstract
Attributed graph clustering aims to group nodes that are similar with respect to both graph structure and node attributes. However, effectively combining these two sources of information remains challenging, as structural proximity does not necessarily imply attribute similarity. The problem becomes even more difficult in high-dimensional settings, where clusters may depend on different subsets of features. To address these issues, we propose PRAGSC, a framework that integrates PageRank-based propagation with subspace clustering. Our propagation scheme incorporates a teleportation mechanism that mitigates over-smoothing by re-injecting the original node features at each iteration, while the subspace clustering component identifies cluster-specific attribute subsets. Experiments on six real-world benchmark datasets demonstrate that PRAGSC consistently outperforms recent state-of-the-art attributed graph clustering methods across different metrics, while enhancing interpretability through the identification of cluster-specific feature subspaces.