CGCFS: A comparative graph clustering method based on feature and structural information

In the domain of deep graph clustering, contrastive learning has garnered significant attention owing to its commendable performance. While Graph Convolutional Networks (GCNs) perform well in certain scenarios, their coupling between filters and weight matrices can lead to over-smoothing, thereby diminishing model performance and robustness. Additionally, inappropriate graph data augmentation, such as diffusion or deleting edges, may disrupt the original structure, leading to semantic drift. To address these issues, this paper introduces a Contrastive Graph Clustering algorithm based on Feature and Structural embedding fusion (CGCFS) that optimizes and improves upon existing methods. Specifically, we replace the GCN with a Multi-Layer Perceptron (MLP), which serves as the backbone of the network. The CGCFS comprises three main modules: Data Preprocessing, which applies low-pass filtering to denoise features while introducing a moderate amount of noise disturbance to the feature samples; Data Augmentation, which employs non-shared parameter MLPs to generate two augmented views of the same node features, enriching representational diversity; and Fusion Contrast, which achieves feature fusion across different augmented views and enhances the network’s discriminative power through a structural consistency objective function. Extensive experiments on seven benchmark datasets demonstrate the significant performance advantages of the CGCFS algorithm.