Adaptive feature boosting and distribution refinement for graph clustering

Deep graph clustering is a common technique for the analysis of large-scale unlabeled graph data, which is fundamental yet challenging. Recently, the combination of Auto-Encoder and Graph Neural Networks in graph clustering methods has gained widespread attention. However, we observe that the weakly relevant features and unreliable target distribution weaken the discriminative representation of graph learning, which limits the graph clustering performance. To tackle this issue, we propose a novel method termed Adaptive feature Boosting and distribution Refinement (ABR) for graph clustering. Specifically, we adaptively enhance the model's capability to capture crucial graph structure and node attributes in each learning iteration through a local-global feature boosting approach. Moreover, we design a dynamic distribution refinement strategy that measures the importance of various soft assignment distributions to generate a robust target distribution. Extensive experiments conducted on five benchmark datasets have demonstrated that our proposed ABR achieves an average improvement of 3.63 % on four metrics compared to the suboptimal graph clustering method.