SCISE framework breaks structural isolation for scalable graph clustering

A new paper accepted to VLDB 2026 presents SCISE (Scalable unsupervised graph Clustering with structural Integrity via community-aware Sampling and Structural Entropy), a framework designed to overcome the structural isolation problem that plagues mini-batch training in graph contrastive learning. Traditional methods lose global community structure when processing graphs in batches, leading to fragmented clusters. SCISE addresses this with three key innovations: the Structural Entropy Community Constraint operator (SECC) that optimizes partition cohesion within a constrained solution space, a Community-Aware Sampling Expansion mechanism (CSampE) that incorporates target node community context into sampling batches, and a Structural Contrastive Learning module (StructCL) that refines edge weights based on intra-batch structural similarity.

Extensive experiments on six mainstream benchmark datasets demonstrate that SCISE significantly outperforms state-of-the-art graph clustering algorithms. The authors also provide ablation studies and robustness analyses confirming the framework's effectiveness for real-world large-scale graphs. The paper is 18 pages with 15 figures and 15 tables, authored by Jingyun Zhang, Hao Peng, Jianxin Li, Angsheng Li, and Philip S. Yu. The work promises to enable more accurate unsupervised clustering of massive networks, with applications in social network analysis, recommendation systems, and biological network mining.