Global Graph Propagation with Hierarchical Information Transfer for Incomplete Contrastive Multi-view Clustering

TL;DR

GHICMC addresses incomplete multi-view clustering by integrating view-specific GCN encoders, a global graph propagation module with hierarchical information transfer across views, and a cluster-level contrastive objective into an end-to-end framework. It computes a consensus representation Z from per-view embeddings and applies stacked graph convolutions on a learnable global graph A to produce deep representations while imputing missing data through cross-view transfers, guided by reconstruction, cross-view consistency, and cluster-level losses. A weight-sharing pseudo-classifier yields per-view clustering assignments Y^v, which are combined to form the final clustering, optimized via a joint objective. Experiments on five datasets demonstrate superior performance over state-of-the-art IMVC methods across ACC, NMI, and ARI, especially at higher missing rates, highlighting the method's robustness and practical impact for incomplete multi-view data.

Abstract

Incomplete multi-view clustering has become one of the important research problems due to the extensive missing multi-view data in the real world. Although the existing methods have made great progress, there are still some problems: 1) most methods cannot effectively mine the information hidden in the missing data; 2) most methods typically divide representation learning and clustering into two separate stages, but this may affect the clustering performance as the clustering results directly depend on the learned representation. To address these problems, we propose a novel incomplete multi-view clustering method with hierarchical information transfer. Firstly, we design the view-specific Graph Convolutional Networks (GCN) to obtain the representation encoding the graph structure, which is then fused into the consensus representation. Secondly, considering that one layer of GCN transfers one-order neighbor node information, the global graph propagation with the consensus representation is proposed to handle the missing data and learn deep representation. Finally, we design a weight-sharing pseudo-classifier with contrastive learning to obtain an end-to-end framework that combines view-specific representation learning, global graph propagation with hierarchical information transfer, and contrastive clustering for joint optimization. Extensive experiments conducted on several commonly-used datasets demonstrate the effectiveness and superiority of our method in comparison with other state-of-the-art approaches. The code is available at https://github.com/KelvinXuu/GHICMC.

Figures (6)

• Figure 1: The overall framework of GHICMC. It consists of three main modules: view-specific representation learning, global graph propagation with hierarchical information transfer, and a contrastive clustering module. Firstly, the view-specific GCN is employed to learn the view-specific representation $\{\bm{\bar{Z}}^v\}^V_{v=1}$ with the local graph, which is then fused to obtain the consensus representation $\bm{Z}$. Secondly, a global graph propagation module with hierarchical information transfer is used to impute the representations adaptively and simultaneously learn deeper representations $\{\bm{H}^v\}^V_{v=1}$ with the consensus representation $\bm{Z}$ and global graph $\bm{A}$ which is constructed by $\{\bm{\bar{A}}^v\}^V_{v=1}$. Finally, the weight-sharing pseudo-classifier is employed to generate the clustering predictions, from which we derive the final clustering result.
• Figure 2: The parameter sensitivity analysis of $\alpha$ and $\beta$ on MSRC-ataset with missing rate of 0.5.
• Figure 3: The influence of parameter $K$ on HandWritten dataset with missing rate of 0.5.
• Figure 4: Comparison of the clustering performance on MSRC-V1 as missing rate increases, with padding representing the standard deviation.
• Figure 5: The visualization results on HandWritten dataset with different missing rates.