Neighborhood Information Aggregation and Multi-View Feature Extraction-Based Contrastive Graph Clustering

In recent years, graph contrastive clustering has received increasing attention in the field of graph deep clustering and achieved very excellent performance. Although graph contrast clustering has shown significant results in this field, most of the existing methods rely on manually designed data enhancement strategies. While these strategies perform well on image data, they often tend to lead to semantic drift when used on graph-structured data, thus limiting the performance of the model. In addition, existing methods mainly rely on the original graph topology information and fail to fully utilize the neighborhood information hidden in the node attribute features. To address the above problems, we proposes a Neighborhood Information Aggregation and Multi-View Feature Extraction-Based Contrastive Graph Clustering (NIA-MVFE-CGC) framework, which improves the existing methods from the perspectives of network architecture, feature redundancy and neighborhood information. First, We directly use multiple multilayer perceptrons (MLPs) to generate multiple views instead of using data augmentation methods. Secondly we utilize mutual information to reduce the redundancy between feature dimensions. Then, we design a neighborhood information aggregation module for mining the neighborhood information relationships of the samples. This module not only considers the explicit structures in the data, but also generates a new neighborhood relationship graph by combining the learned potential relationship structures. In addition, we design a weight graph that allows the model to adaptively adjust the proximity between samples during the learning process. Extensive experiments on five benchmark datasets show that our proposed method outperforms most other clustering algorithms.