Deep Incomplete Multi-View Clustering via Hierarchical Imputation and Alignment

TL;DR

DIMVC-HIA tackles incomplete multi-view clustering by unifying hierarchical imputation with dual alignment: an energy-based semantic alignment to enforce intra-cluster compactness and a contrastive assignment alignment to harmonize cross-view predictions. It first imputes missing cluster assignments using cross-view semantic similarity, then recovers missing latent features via cluster-aware prototypes, all within an end-to-end framework built on view-specific autoencoders and a shared clustering predictor. The approach demonstrates strong, stable performance across diverse datasets and missingness levels, outperforming state-of-the-art IMVC methods and showing notable gains as incompleteness increases. This work advances practical IMVC by mitigating error propagation in imputation and reducing representation uncertainty through structured semantic alignment.

Abstract

Incomplete multi-view clustering (IMVC) aims to discover shared cluster structures from multi-view data with partial observations. The core challenges lie in accurately imputing missing views without introducing bias, while maintaining semantic consistency across views and compactness within clusters. To address these challenges, we propose DIMVC-HIA, a novel deep IMVC framework that integrates hierarchical imputation and alignment with four key components: (1) view-specific autoencoders for latent feature extraction, coupled with a view-shared clustering predictor to produce soft cluster assignments; (2) a hierarchical imputation module that first estimates missing cluster assignments based on cross-view contrastive similarity, and then reconstructs missing features using intra-view, intra-cluster statistics; (3) an energy-based semantic alignment module, which promotes intra-cluster compactness by minimizing energy variance around low-energy cluster anchors; and (4) a contrastive assignment alignment module, which enhances cross-view consistency and encourages confident, well-separated cluster predictions. Experiments on benchmarks demonstrate that our framework achieves superior performance under varying levels of missingness.

Figures (5)

• Figure 1: The overall architecture of our proposed DIMVC-HIA framework.
• Figure 2: Ablation study results on the Fashion-MNIST dataset with $\eta$ = 0.5.
• Figure 3: T-SNE visualizations of embeddings on BDGP with $\eta = 0.5$ (left) and Handwritten with $\eta = 0.7$ (right).
• Figure 4: Convergence curves for the MNIST-USPS and Fashion-MNIST datasets with $\eta = 0.1$.
• Figure 5: Parameter sensitivity of $\alpha$ and $\beta$ on BDGP (left) and Handwritten (right) datasets with $\eta = 0.3$, where both $\alpha$ and $\beta$ take values in $\{0.001, 0.01, 0.05, 0.1, 1\}$.