Abstract
Multi-view clustering aims to integrate information from multiple sources to improve clustering performance. However, practical applications often encounter incomplete data where some views are missing for certain samples. This paper introduces the Unified Embedding Alignment Framework (UEAF), a novel approach that simultaneously recovers missing views and learns a consensus latent representation. By incorporating reverse graph regularization and adaptive view weighting, UEAF preserves local geometric structures and handles varying view quality. Experimental results across multiple datasets demonstrate that UEAF significantly outperforms state-of-the-art incomplete multi-view clustering methods.
Introduction
Multi-view clustering (MVC) leverages complementary information from different data sources—such as text, images, and audio—to achieve more robust clustering than single-view methods. A significant hurdle in real-world MVC is the presence of incomplete data, where some samples lack one or more views. Traditional incomplete multi-view clustering (IMVC) methods often struggle by either discarding incomplete samples, filling gaps with simple averages, or failing to capture the underlying geometric structure of the data.
The Incomplete Multi-View Clustering Problem
Existing IMVC approaches typically suffer from three primary limitations:
- Loss of Semantic Information: Discarding or crudely filling missing views ignores the latent relationships between available and missing features.
- Neglect of Local Geometry: Many models fail to ensure that data points which are neighbors in the original feature space remain neighbors in the learned consensus space.
- Uniform View Weighting: Treating all views as equally important is suboptimal, especially when views have varying levels of completeness or noise.
The Unified Embedding Alignment Framework (UEAF)
The proposed UEAF addresses these challenges by unifying missing view inference and representation learning into a single optimization framework.
Core Components
- Missing View Inference: UEAF introduces an error matrix to model missing data. It employs Laplacian regularization to ensure that the inferred missing views are consistent with the known features of available data.
- Consensus Representation Learning: The framework learns a common latent matrix $P$ that aligns both available and recovered views into a shared semantic space.
- Reverse Graph Regularization: To preserve local structure, UEAF learns a similarity graph and imposes a rank constraint on its Laplacian matrix. This ensures the graph has exactly $c$ connected components, directly aligning the feature learning with the clustering objective.
- Adaptive Weighting: The model automatically assigns weights to different views based on their informativeness and completeness, enhancing the influence of high-quality data.
Experimental Evaluation
The authors evaluated UEAF on four diverse datasets: BUAA-visnir (face images), Handwritten Digits, 3 Sources (news), and BBCSport. Incompleteness was simulated by randomly removing views at rates between 10% and 70%.
Results and Discussion
UEAF consistently outperformed eight state-of-the-art methods, including PVC, IMG, and DAIMC.
Key findings include:
- Robustness to Incompleteness: On the BUAA dataset, UEAF maintained high performance even with very low paired rates. For instance, at a 0.1 paired rate, UEAF achieved an accuracy (ACC) of approximately 73%, significantly outperforming degraded models and competing methods.
- Ablation Study: Component analysis revealed that both missing view inference and graph regularization are critical; removing either component led to a sharp drop in performance, confirming the necessity of the "unified" approach.
- Parameter Sensitivity: While robust, the model involves hyperparameters ($\lambda_1, \lambda_2, \lambda_3$) that require tuning for optimal results.
Conclusion
UEAF provides a comprehensive solution for incomplete multi-view clustering by intelligently filling data gaps while learning a structured consensus representation. Future work may focus on reducing the $O(n^3)$ computational complexity to improve scalability and exploring deep learning architectures to capture non-linear data relationships.
Summary updated