Imputation-free and Alignment-free: Incomplete Multi-view Clustering Driven by Consensus Semantic Learning

TL;DR

IMMVC suffers from missing data causing prototype shifts and cross-view misalignment. FreeCSL provides an imputation- and alignment-free solution by learning a shared semantic space with consensus prototypes $igoldsymbol{C}=igigoldsymbol{c}_kigig|_{k=1}^K$, using prototypical protocontrast and a modularity-driven graph-clustering module to fuse cross-view and within-view information. The method optimizes a joint objective $\mathcal{L}=\mathcal{L}_{rec}+\mathcal{L}_{cc}+\\mathcal{L}_{gc}$ across reconstruction, cross-view consensus semantic learning, and within-view cluster enhancement, then performs $k$-means on the consensus representations. Empirical results on four IMVC benchmarks show FreeCSL achieves superior accuracy and robustness, especially under high missing rates and large-scale settings, demonstrating reliable clustering without imputation or alignment.

Abstract

In incomplete multi-view clustering (IMVC), missing data induce prototype shifts within views and semantic inconsistencies across views. A feasible solution is to explore cross-view consistency in paired complete observations, further imputing and aligning the similarity relationships inherently shared across views. Nevertheless, existing methods are constrained by two-tiered limitations: (1) Neither instance- nor cluster-level consistency learning construct a semantic space shared across views to learn consensus semantics. The former enforces cross-view instances alignment, and wrongly regards unpaired observations with semantic consistency as negative pairs; the latter focuses on cross-view cluster counterparts while coarsely handling fine-grained intra-cluster relationships within views. (2) Excessive reliance on consistency results in unreliable imputation and alignment without incorporating view-specific cluster information. Thus, we propose an IMVC framework, imputation- and alignment-free for consensus semantics learning (FreeCSL). To bridge semantic gaps across all observations, we learn consensus prototypes from available data to discover a shared space, where semantically similar observations are pulled closer for consensus semantics learning. To capture semantic relationships within specific views, we design a heuristic graph clustering based on modularity to recover cluster structure with intra-cluster compactness and inter-cluster separation for cluster semantics enhancement. Extensive experiments demonstrate, compared to state-of-the-art competitors, FreeCSL achieves more confident and robust assignments on IMVC task.

Figures (18)

• Figure 1: Research Motivation for Consensus Semantic Learning. (a) two existing paradigms, instance- and cluster-level, either treat cross-view unpaired observations with similar semantics as false negatives or neglect to treat within-view observations with similar semantics as true positives; (b) we propose a novel semantic-level paradigm based on contrastive clustering with a set of consensus prototypes to foster semantic consistency across all view data.
• Figure 2: The framework of FreeCSL. (a) Reconstruction module, encodes observations into clustering-friendly representations for each view; (b) Consensus semantic learning module, learns semantic representations through cross-view contrastive clustering in a shared semantic space, where paired observations are assigned to their nearest semantic prototype for consistent assignments. (c) Cluster semantic enhancement module, enriches semantic representations with cluster structure information through within-view graph clustering, which applies GCN to aggregate view-specific semantic information and maximizes spectral modularity to recover cluster structure with greater separation and lower entropy. Ultimately, perform $k$-means on the consensus semantic representations to predict cluster labels.
• Figure 3: Visualization for Table \ref{['tab:performance']} based on metric ACC.
• Figure 4: Similarity matrices of $\{\mathbf{Z}^{{v}}\}_{v=1}^{4}$, $\mathbf{Z}$ without consensus semantic learning on ALOI-100 with $r=0.5$.
• Figure 5: Similarity matrices of $\{\mathbf{H}^{{v}}\}_{v=1}^{4}$, $\mathbf{H}$ with consensus semantic learning on ALOI-100 with $r=0.5$.

Theorems & Definitions (6)

• Theorem 1
• Theorem 2
• Theorem 3
• proof
• Theorem 4
• proof