Multi-Scale Diffusion Enhancement for Graph Clustering with Heat and PPR Kernels

Abstract

In context of graph clustering, existing contrastive learning approaches relying on adjacency and diffusion matrices often fail to capture the complex structural patterns (long range dependency and cluster hierarchies) across different scales − hops from a node. In thiswork, we propose a multi-scale diffusion enhancement policy using the Personalized PageRank (PPR) kernel and a Heat kernel to address this challenge. Our proposed approach uses complementary diffusion process that capture the clustering patterns to overcome the shortcomings of conventional single-scale diffusion and Conventional k-hop multi-scale diffusion. We carried out necessary performance comparisons in our experiments against multiple benchmarks, and found notable gains, thereby confirming the effectiveness of multi-scale kernel integration.