Deep Incomplete Multi-view Clustering with Distribution Dual-Consistency Recovery Guidance

TL;DR

This work tackles incomplete multi-view clustering by addressing inter-view heterogeneity during recovery. It introduces BURG, a framework that performs cross-view distribution transfer using flow-based models and enforces dual-consistency—neighbor-aware and prototypical—to guide both recovery and representation learning. The method combines three components: multi-view feature extraction, distribution transfer learning, and dual-consistency guided recovery, and demonstrates state-of-the-art clustering performance across six benchmarks with varying missing rates. The results highlight BURG's ability to restore realistic missing-view representations and enhance inter-view clustering structure, offering scalable and robust performance for practical multi-view data scenarios.

Abstract

Multi-view clustering leverages complementary representations from diverse sources to enhance performance. However, real-world data often suffer incomplete cases due to factors like privacy concerns and device malfunctions. A key challenge is effectively utilizing available instances to recover missing views. Existing methods frequently overlook the heterogeneity among views during recovery, leading to significant distribution discrepancies between recovered and true data. Additionally, many approaches focus on cross-view correlations, neglecting insights from intra-view reliable structure and cross-view clustering structure. To address these issues, we propose BURG, a novel method for incomplete multi-view clustering with distriBution dUal-consistency Recovery Guidance. We treat each sample as a distinct category and perform cross-view distribution transfer to predict the distribution space of missing views. To compensate for the lack of reliable category information, we design a dual-consistency guided recovery strategy that includes intra-view alignment guided by neighbor-aware consistency and cross-view alignment guided by prototypical consistency. Extensive experiments on benchmarks demonstrate the superiority of BURG in the incomplete multi-view scenario.

Figures (5)

• Figure 1: Basic idea of cross-view distribution transfer recovery. The consistent distribution obtained from the complete views is transferred to the missing views using view-specific flow models. The recovered representation is then input into the view-specific decoder to generate the recovered data.
• Figure 2: The framework of BURG. We use three views as an example, where the first and second views are complete, while the third is missing. As shown, BURG consists of a joint optimization of three modules: multi-view feature extraction(MFE), distribution transfer learning(DTL), and dual-consistency guided recovery(DGR). The two types of consistency are neighbor-aware and prototypical consistency. In DTL, the common embedding $\mathbf{H}$ is formed by combining $\mathbf{H}^1$ and $\mathbf{H}^2$, which are obtained through view-specific forward flows. Subsequently, $\mathbf{H}$ is passed through reverse flow $(F^3)^{-1}$ to generate the latent representation $\tilde{\mathbf{Z}}^3$ of the missing view. Furthermore, DGR provides essential local intra-view structure and global inter-view clustering structure throughout the training process to ensure the discriminability of the recovered representation.
• Figure 3:
• Figure 4: Visualization of the discrepancy between recovered and ground-truth data on the HandWritten dataset before and after training, with a missing rate of 0.5.
• Figure 5: Sensitivity analysis of $\alpha$ and $\beta$ for our method over HandWritten and Animal (The missing rate is 0.5).