Abstract
Multi-view graph clustering aims to group nodes by integrating information from multiple perspectives. However, existing Graph Convolutional Network (GCN) approaches often fall victim to the "curse of locality," where local smoothing across inter-class edges degrades clustering performance. This paper introduces Consistent Context-Aware Representation Learning (CCARL), which leverages global anchors and Fused Gromov-Wasserstein Optimal Transport to maintain a consistent global manifold structure across views. Experimental results on four real-world datasets demonstrate that CCARL significantly outperforms existing state-of-the-art methods in both accuracy and cluster quality.
Introduction
Graph clustering is a fundamental task in data mining, involving the grouping of nodes based on their features and connectivity. In many real-world scenarios, data is multi-view in nature—for instance, a social network might have one view representing friendships and another representing shared interests. While Graph Convolutional Networks (GCNs) have become the standard for this task, they rely heavily on Laplacian filters that smooth node representations based on immediate neighbors. This leads to the "curse of local manifold": when a graph contains many "inter-class edges" (connections between nodes of different groups), local smoothing makes distinct nodes look similar, hindering clustering performance. To overcome this, a global perspective is required to maintain consistency across all views.
Methodology: CCARL
The proposed CCARL framework addresses the limitations of local smoothing by introducing a global coordinate system. The architecture is centered around three key innovations:
1. Global Anchors and Fused GW-OT
CCARL defines a set of "anchors" within a global space that serve as a fixed reference point. To map these anchors to specific nodes in each view, the model employs Fused Gromov-Wasserstein Optimal Transport (GW-OT). Unlike standard optimal transport, Fused GW-OT simultaneously considers node features and graph topology, ensuring the global manifold is accurately aligned with the local data in each view.
2. Attention-Based Fusion and Context-Aware Projection
Since different views contribute varying levels of information, CCARL uses an attention mechanism to adaptively fuse embeddings. This allows the model to prioritize the most informative views for each node. The fused representation is then projected back onto the global anchors ($Z_c = ZP^T$), transforming local features into a context-aware global coordinate system.
3. Self-Training Strategy
To refine cluster boundaries, CCARL implements a self-training loss. By calculating a target distribution and minimizing the KL-divergence between current predictions and this target, the model effectively tightens clusters around their centroids, enhancing discriminative power.
Experiments and Results
The performance of CCARL was evaluated on four diverse datasets: AMINER (academic), BDGP (genetic), IMDB (movies), and YELP (business).
Quantitative Performance
CCARL consistently outperformed 15 baseline methods. Notably, on the AMINER dataset, CCARL achieved an accuracy (ACC) of 73.96%, a significant improvement over the next best GCN-based method, LMGEC (61.76%). On the BDGP dataset, the model reached 91.08% ACC and 81.08% NMI, demonstrating its robustness across different domains.
Qualitative Visualization
The superiority of CCARL is further evidenced by t-SNE visualizations. As shown in the comparison below, CCARL produces much tighter and more distinct clusters with clearer boundaries than competing methods like MvAGC or CMGEC.
Conclusion
CCARL effectively mitigates the "curse of locality" in multi-view graph clustering by establishing a consistent global context. By using anchors and Fused GW-OT, the model preserves global manifold structures that local GCN filters often obscure. Future work will focus on extending this framework to handle incomplete data scenarios where certain nodes or views may be missing.
Summary updated