Work

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

Year: 2024

Type: article

Abstract: In the domain of deep graph clustering, contrastive learning has garnered significant attention owing to its commendable performance. Nonetheless, it has been observed that while graph convolution operations yield certain outcomes, they are time-intensive and diminish efficiency. Furthermore, the methodology for constructing positive and negative samples in graph clustering, though innovative, suffers from inadequate reliability in negative sample construction, thus impacting network stability. To mitigate these challenges, this paper introduces the Contrastive Graph Clustering Algorithm Based on Feature and Structural Information (CGCFS) to realize optimizations and advancements over extant methodologies, encompassing network architecture, data augmentation techniques, and loss functions. Specifically, CGCFS comprises three key components: data preprocessing, data augmentation, and fusion contrast. The proposed model employs solely a multilayer perceptron (MLP) as the backbone network. Data preprocessing primarily encompasses Laplacian low-pass denoising of features and the introduction of noise-perturbed feature samples. In the data augmentation segment, non-shared parameter MLPs are utilized to generate two enhanced views of identical node information. In the fusion contrast phase, fusion between the feature embedding and structural embedding correlation matrices is achieved, employing a cross-view structural consistency objective function to bolster the representation capability of the embedding space, thereby enhancing the network's discriminative prowess. Extensive experiments conducted on six benchmark datasets corroborate the efficacy and superiority of the proposed algorithm. Relative to the second-best method, our findings exhibit an average enhancement of 0.96% across each evaluation metric (ACC, NMI, ARI, and F1) across the six datasets.

Source: Expert Systems with Applications

DOI: 10.1016/j.eswa.2024.126346

Authors: Xiaoxuan Wu, Mengnan Zhou, Weizhen Zhang

Institutions: Anhui Jianzhu University, Hefei University of Technology, Hefei University

Language: English

Topic: Advanced Graph Neural Networks

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $3,320

Funder: Anhui Provincial Department of Education