Title: Adaptive Local-Global Integration for Robust Contrastive Graph Clustering

Graph clustering plays a pivotal role in graph analysis by uncovering structural patterns and identifying node communities. While recent progress in self-supervised contrastive learning has enhanced clustering performance through the utilization of structural and attribute signals, current approaches face significant challenges. Specifically, they often fail to flexibly capture high-order local structures and tend to neglect global semantics within complex graphs. These shortcomings result in suboptimal node representations, particularly when dealing with real-world graphs characterized by fragmented structures and indistinct cluster boundaries.

To overcome these issues, we propose a novel contrastive graph clustering framework that jointly integrates multi-scale local structures with global semantics using attention mechanisms. At the local level, the model extracts topological signals from multiple propagation depths via Graph Neural Networks (GNNs). These signals are adaptively fused through attention-based weighting to effectively capture multi-scale neighborhood features. Concurrently, at the global level, semantic prototypes—derived from dynamically evolving cluster centers—are adaptively aggregated via attention to guide node representations and improve inter-cluster separability.

The proposed model is trained within a dual-view contrastive learning paradigm. It employs a hybrid objective function that combines instance-level and structure-aware losses, thereby enhancing both the robustness and discriminative power of the representations. Experimental results on eight real-world graph datasets indicate that our method delivers competitive clustering performance. The source code is publicly available at https://github.com/vege12138/w2.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC