A scalable distributed deep graph clustering framework via heterogeneity-aware graph neural networks and contrastive-prototypical learning

The rapid growth of large-scale heterogeneous networks has created significant challenges for existing graph clustering methods, as traditional approaches fail to capture multi-relational semantics, and centralized deep models suffer from scalability and communication constraints. Moreover, the distributed nature of graph-structured data introduces additional challenges related to representation alignment, consistency, and efficient learning across partitions. These limitations highlight the need for a clustering framework that is both semantically expressive and scalable in distributed environments. Despite recent advances in graph neural networks and self-supervised learning, current methods either struggle to maintain consistent embeddings across distributed subgraphs or rely on handcrafted augmentations that lack generalization across heterogeneous modalities. In this paper, a Distributed Deep Graph Clustering (DDGC) framework is proposed, integrating heterogeneity-aware graph neural networks with a contrastive-prototypical learning objective. The framework processes partitioned heterogeneous graphs, generates semantically aligned node embeddings via multi-view encoders, and refines clusters through distributed contrastive optimization. The proposed method is evaluated on real-world heterogeneous graph datasets using clustering accuracy (ACC), normalized mutual information (NMI), adjusted Rand index (ARI), silhouette score, communication overhead, and runtime as evaluation metrics. Experimental results demonstrate that the proposed framework improves clustering accuracy by approximately 9.2%, NMI by 8.5%, and ARI by 7.8%, while enhancing the silhouette score by 6.9% compared to baseline methods. In addition, communication overhead is reduced by 10.6%, and runtime efficiency is improved by 8.1%. These results confirm that the proposed approach achieves superior clustering performance, better cluster separability, and enhanced scalability, making it a practical solution for large-scale distributed heterogeneous network analysis.