[Paper Review] Dual Information Enhanced Multi-view Attributed Graph Clustering

Jiaqi Lin, Man-Sheng Chen+3 more|arXiv (Cornell University)|Nov 28, 2022

Advanced Graph Neural Networks4 citations

TL;DR

This paper proposes DIAGC, a novel multi-view attributed graph clustering method that disentangles consensus and specific information via a Specific Information Reconstruction (SIR) module and enhances high-level representation learning through Mutual Information Maximization (MIM), achieving state-of-the-art performance on real-world benchmarks by effectively capturing both low-level and high-level graph structures and attributes.

ABSTRACT

Multi-view attributed graph clustering is an important approach to partition multi-view data based on the attribute feature and adjacent matrices from different views. Some attempts have been made in utilizing Graph Neural Network (GNN), which have achieved promising clustering performance. Despite this, few of them pay attention to the inherent specific information embedded in multiple views. Meanwhile, they are incapable of recovering the latent high-level representation from the low-level ones, greatly limiting the downstream clustering performance. To fill these gaps, a novel Dual Information enhanced multi-view Attributed Graph Clustering (DIAGC) method is proposed in this paper. Specifically, the proposed method introduces the Specific Information Reconstruction (SIR) module to disentangle the explorations of the consensus and specific information from multiple views, which enables GCN to capture the more essential low-level representations. Besides, the Mutual Information Maximization (MIM) module maximizes the agreement between the latent high-level representation and low-level ones, and enables the high-level representation to satisfy the desired clustering structure with the help of the Self-supervised Clustering (SC) module. Extensive experiments on several real-world benchmarks demonstrate the effectiveness of the proposed DIAGC method compared with the state-of-the-art baselines.

Motivation & Objective

• To address the limitation of existing methods in ignoring view-specific structural information during consensus representation learning.
• To improve clustering performance by recovering meaningful high-level representations from low-level features in multi-view attributed graphs.
• To jointly optimize consensus learning and view-specific information disentanglement in a unified deep clustering framework.
• To enhance representation quality by maximizing mutual information between low-level and high-level representations.
• To validate the effectiveness of each component through ablation studies and comprehensive benchmark evaluation.

Proposed method

• The Specific Information Reconstruction (SIR) module disentangles consensus and specific information by reconstructing view-specific adjacency matrices from low-level representations, enabling GCN to learn more essential features.
• The Mutual Information Maximization (MIM) module maximizes agreement between low-level representations and a shared high-level representation, preserving clustering-oriented structure.
• A self-supervised clustering (SC) module enforces the high-level representation to align with desired cluster assignments using soft labels and KL divergence.
• The overall objective function combines reconstruction loss, mutual information maximization, and clustering loss to jointly optimize all components.
• The model is trained end-to-end using an alternating optimization strategy, updating the encoder, clustering centroids, and representation distributions iteratively.
• t-SNE visualization is used to qualitatively validate the clarity and separability of learned representations.

Experimental results

Research questions

• RQ1 How does disentangling consensus and specific information improve clustering performance in multi-view attributed graphs?
• RQ2 To what extent does mutual information maximization between low-level and high-level representations enhance clustering structure learning?
• RQ3 How robust is the proposed DIAGC method to hyperparameter variations across diverse real-world datasets?
• RQ4 What is the individual contribution of each component (SIR, MIM, SC) to the overall clustering performance?
• RQ5 Can the proposed method achieve better clustering results than state-of-the-art baselines on multi-view attributed graph benchmarks?

Key findings

• The ablation study shows that removing the MIM module causes a significant drop in clustering performance, indicating its critical role in preserving high-level structure.
• Without the SIR module, clustering performance degrades due to interference from view-specific information in low-level representations.
• The SC module improves NMI by 5.58%, 1.03%, and 0.97% on ACM, DBLP, and IMDB respectively, confirming its effectiveness in refining cluster assignments.
• The method demonstrates robustness to hyperparameter α across all benchmarks, with stable performance across the tested range.
• t-SNE visualization confirms that DIAGC produces clearer and more separable cluster structures compared to baselines like DEAGC and O2MAC.
• Convergence analysis shows that the loss function monotonically decreases over iterations, indicating stable training dynamics.