MoEGCL: Mixture of Ego-Graphs Contrastive Representation Learning for Multi-View Clustering

TL;DR

MoEGCL tackles coarse-grained graph fusion in multi-view clustering by fusing ego graphs at the sample level with a Mixture-of-Experts framework (MoEGF) and by enforcing cluster-aware contrastive learning (EGCL). The approach combines per-view autoencoders, ego-graph construction, gated fusion across views, and a clustering objective to achieve fine-grained, semantically coherent representations. Empirical results on six public datasets show state-of-the-art performance, supported by comprehensive ablations, visualization, and convergence analyses. The work highlights the practical value of sample-level ego-graph fusion for improved deep MVC performance and representation quality.

Abstract

In recent years, the advancement of Graph Neural Networks (GNNs) has significantly propelled progress in Multi-View Clustering (MVC). However, existing methods face the problem of coarse-grained graph fusion. Specifically, current approaches typically generate a separate graph structure for each view and then perform weighted fusion of graph structures at the view level, which is a relatively rough strategy. To address this limitation, we present a novel Mixture of Ego-Graphs Contrastive Representation Learning (MoEGCL). It mainly consists of two modules. In particular, we propose an innovative Mixture of Ego-Graphs Fusion (MoEGF), which constructs ego graphs and utilizes a Mixture-of-Experts network to implement fine-grained fusion of ego graphs at the sample level, rather than the conventional view-level fusion. Additionally, we present the Ego Graph Contrastive Learning (EGCL) module to align the fused representation with the view-specific representation. The EGCL module enhances the representation similarity of samples from the same cluster, not merely from the same sample, further boosting fine-grained graph representation. Extensive experiments demonstrate that MoEGCL achieves state-of-the-art results in deep multi-view clustering tasks. The source code is publicly available at https://github.com/HackerHyper/MoEGCL.

Figures (5)

• Figure 1: Mixture of Ego-Graphs Fusion. Firstly, an ego graph for each sample in each view is constructed, and input each ego graph as an expert into the Mixture-of-Experts (MoE) architecture and calculate the importance of each expert. Then, the importance coefficient with the experts are merged to form a fused ego graph. Lastly, aggregate the ego graphs of all samples to generate a global connected graph.
• Figure 2: The architecture diagram of MoEGCL. The MoEGCL method contains two modules: MoEGF and EGCL. MoEGF first constructs ego graphs and then aggregates ego graphs of all views into a fused ego graph by a Mixture-of-Experts network. In addition, the EGCL module is proposed to improve the representation similarity of samples in the same cluster, as opposed to only concentrating on the consistency at the sample level. It significantly enhances fine-grained graph representation.
• Figure 3: The visualization results of the fused representations $\{\hat{h}_{i}\}^N_{i=1}$ on the LGG, MNIST, and Caltech5V datasets after convergence. From the visualization results of the three subfigures, it can be concluded that our proposed method separates the samples completely in the feature space, thus proving that our method is very effective.
• Figure 4: The convergence analysis on the Caltech5V dataset. In the figure, the test ACC, NMI, and PUR are shown at the top, and the training loss is depicted at the bottom. It can be observed that around $400$ epochs, MoEGCL reaches a steady state, with the training loss no longer decreasing, and achieves the best accuracy compared with SOTA methods.
• Figure 5: The parameter analysis on the Caltech5V dataset. The figure shows the changes in three evaluation metrics: ACC, NMI, and PUR. The metrics are influenced by two hyperparameters $\lambda$ and $\tau$. $\lambda$ is the combination coefficient of two loss functions. $\tau$ denotes the temperature coefficient.