Introduction
Multi-view data, which integrates information from diverse sources (e.g., medical images and patient history), is prevalent in real-world applications. However, data incompleteness—where certain views are missing for some samples—poses a significant challenge for clustering. This paper addresses Incomplete Multi-view Clustering (IMVC) by proposing a novel framework: Joint View Imputation and Label Generation (JVILG).
Problem Statement
Existing anchor-based methods for IMVC often suffer from two primary limitations. First, they typically employ a two-stage process—constructing an anchor graph followed by spectral clustering—which can lead to information loss and suboptimal results. Second, their imputation strategies for missing views often rely on noisy or redundant data from other samples, failing to leverage the representative power of anchors effectively.
Proposed Method: JVILG
JVILG treats clustering and data recovery as a unified reconstruction problem, utilizing "anchors" as both representatives for clustering and tools for imputation. The framework consists of two core modules:
- Anchor-based tensorized Label Generation (ALG): This module directly links raw data, fine-grained anchors, and cluster labels via a membership matrix. To ensure cross-view consistency, it employs tensor Schatten p-norm regularization, treating labels from all views as a 3D tensor to capture complex spatial correlations.
- Anchor-based sparse regularized Subspace Correlation (ASC): This module reconstructs missing views as sparse linear combinations of anchors from the same latent subspace. $L_1$-norm regularization ensures that only the most relevant anchors are used, minimizing the introduction of noise during imputation.
The entire system is optimized using an alternating algorithm based on the Augmented Lagrangian Multiplier (ADMM) method, ensuring convergence to a stable solution.
Experimental Evaluation
The performance of JVILG was evaluated on six benchmark datasets: ProteinFold, Caltech101-7, BDGP, CCV, Animal, and NUSWIDEOBJ. The experiments covered missing data ratios ranging from 10% to 90%.
Key Results and Discussion
JVILG consistently outperformed both traditional and deep learning-based state-of-the-art methods across all metrics, including Accuracy (ACC), Normalized Mutual Information (NMI), and Purity.
As shown in the figure above, JVILG (represented by the black line) maintains high performance and remarkable stability even as the missing ratio increases to 90%. For instance, on the CCV dataset at a 10% missing ratio, JVILG achieved a 13.73% improvement in Accuracy over the second-best method, DIVIDE. The model's robustness under extreme data loss highlights the effectiveness of its joint imputation and label generation strategy.
Conclusion and Future Work
JVILG demonstrates that anchors can be effectively utilized for both data imputation and direct label generation, bypassing the limitations of graph-based two-stage methods. While the model shows strong performance, future research could explore more dynamic anchor selection processes and optimize computational efficiency for large-scale datasets.
Summary updated