Rectified Attribute-Missing Graph Clustering

Qianqian Wang, Xiang Long, Wei Feng, Bin Liu

Keywords: Attribute-Missing Graphs, Deep Graph Clustering, Unsupervised Learning

TL;DR: a novel end-to-end unified framework conclusively for attribute-missing graph clustering

Abstract:

Deep Graph Clustering(DGC) has gained widespread attention in some tasks like face recognition and social network analysis because of its powerful capability of capturing the latent distribution of multi-view graph-structured data and grouping nodes in the graph into different clusters. However, in the real world, they often face the situation where attributes of some nodes are missing, resulting in clustering performance degradation. Although many methods have been developed to mitigate the problem, most of them are two-stage methods that separate embedding learning from clustering, and may cause deviation of node embedding because of the missing nodes.
To address this problem, we propose a novel end-to-end attribute-missing graph clustering learning method termed $\underline{R}$ectified $\underline{A}$ttribute-$\underline{M}$issing $\underline{G}$raph $\underline{C}$lustering (R-AMGC). First, it performs two augmentations to generate two attribute views, and the missing attributes are set to be learnable in one of the views. Subsequently, we maximize the mutual information between two encoded views via contrastive learning and then effectively mitigate the embedding distortion. Additionally, to learn clustering-friendly embedding, we design a module termed triple constraint, which not only maintains the alignment between graph structure and cluster assignment but also captures structural and attribute information. Extensive experiments on four graph datasets have strongly validated the effectiveness and superiority of R-AMGC compared to other counterparts.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 8350