Abstract
Multi-view clustering (MVC) and multi-view graph clustering (MVGC) have traditionally been treated as distinct tasks. This paper presents GMVC (Graph Embedded Contrastive Learning for Multi-View Clustering), a unified framework designed to bridge this gap. GMVC utilizes adaptive graph convolutions to unify data formats and introduces a graph-guided contrastive loss to mitigate the "false associations" problem—where semantically similar samples are incorrectly treated as negatives. Experimental results across eight benchmarks demonstrate that GMVC achieves state-of-the-art performance in both general and graph-based multi-view scenarios.
Introduction
Multi-view clustering aims to integrate information from diverse data sources (views) to improve grouping accuracy. However, existing methods often specialize in either feature-based MVC or structure-based MVGC, lacking a common methodology. Furthermore, standard contrastive learning approaches often suffer from "false associations," where samples belonging to the same cluster are pushed apart because they are treated as negative pairs. GMVC addresses these challenges by constructing a unified graph-based representation and using a learned global graph to guide the contrastive learning process.
The GMVC Framework
The GMVC architecture consists of three primary components: data unification, view-specific pre-training, and cross-view contrastive learning.
1. Data Format Unification
To handle both feature vectors and pre-existing graphs, GMVC employs a self-expression mechanism to generate adjacency matrices for general multi-view data. This ensures a consistent input format across different data types.
2. Adaptive Graph Convolution (AGC)
Within each view, an AGC module aggregates information from multi-order neighbors. The weights for different orders are learned adaptively to filter noise and emphasize relevant structural information. An autoencoder then maps these features into a latent space $Z$.
3. Graph-Guided and Cluster-Guided Learning
The core of GMVC lies in its dual contrastive learning strategy:
- Graph-Guided Contrastive Learning (GGC): An attention mechanism aggregates view-specific graphs into a global graph. This global graph identifies reliable neighbors, ensuring that similar samples are not penalized by the contrastive loss.
- Cluster-Guided Contrastive Clustering (CGC): This module aligns cluster distributions across views, ensuring consistency in the final assignments.
Experimental Results
GMVC was evaluated on eight datasets, including DHA, NGs, WebKB, and ACM. The model demonstrated significant improvements over baselines:
- WebKB: Achieved 80.2% accuracy, an 18.5% improvement over K-Means.
- ACM: Reached 93.4% accuracy, outperforming VGAE by 49%.
Ablation studies confirmed that both the AGC and GGC components are critical for performance, as removing either led to substantial accuracy drops.
Qualitative Analysis
The effectiveness of the learning process is visualized through t-SNE embeddings and adjacency matrix reconstructions over training epochs.
As shown in the figure, the clusters become significantly more distinct by Epoch 50 compared to the initial state. The learned adjacency matrices (c and f) exhibit clear block-diagonal structures, indicating that the model successfully captures the underlying cluster relationships.
Conclusion
GMVC provides a robust, unified framework for multi-view clustering by integrating graph embedding with guided contrastive learning. By leveraging a global graph to supervise contrastive pairs, it effectively overcomes the "false association" problem. Future work will focus on improving scalability for large-scale datasets and handling incomplete views where some modalities may be missing.
Summary updated