Introduction
Multi-view clustering (MVC) is a fundamental task in machine learning that seeks to group data points by integrating information from multiple sources or "views." For instance, an autonomous vehicle might process data from cameras, LiDAR, and radar simultaneously to identify objects. However, real-world applications often suffer from the problem of missing views due to sensor failures or data corruption, leading to the challenge of Incomplete Multi-view Clustering (IMVC).
Existing IMVC methods typically rely on imputation—filling in missing data—but they often struggle with two primary issues: high data requirements for training recovery modules and the production of "blurry" or non-discriminative samples that fail to aid clustering. To address these challenges, this paper introduces Diffusion Contrastive Generation (DCG), a framework that leverages the generative power of diffusion models and the discriminative nature of contrastive learning to recover missing views and enhance cluster separation.
The DCG Framework
The DCG architecture is composed of four synergistic modules designed to handle feature extraction, data recovery, and cluster assignment:
- View-specific Feature Learning: The model utilizes deep autoencoders to compress available views into low-dimensional latent representations, ensuring that the essential characteristics of each view are captured.
- Diffusion Contrastive Generation: This is the core of the framework. Unlike standard imputation, DCG employs a Diffusion Model to learn the underlying data distribution through iterative forward noise-addition and reverse denoising. By incorporating contrastive learning on a limited set of paired samples, the model aligns generated views with real data, ensuring high-quality recovery even with minimal complete data.
- Instance-level Interactive Learning: An attention-based fusion mechanism dynamically weights the importance of different views for each sample. This module utilizes Maximum Mutual Information to maintain consistency between the fused representation and the individual views.
- Category-level Interactive Learning: A shared classifier generates "soft" cluster labels. By applying category-level contrastive loss and KL divergence, the model sharpens cluster boundaries, leading to more distinct and separable groups.
Diffusion as a Cluster Tightener
A significant insight of this research is the dual role of the diffusion process. Beyond simple data recovery, the reverse diffusion (denoising) process acts as a natural clustering mechanism. As the model removes noise, representations are not merely restored to their original state; they are actively pushed toward the "prototype" or center of their respective categories.
Furthermore, the authors utilize the sinusoidal positional encoding of the diffusion time steps to perform inference-time extrapolation. By running more denoising steps during inference ($T_{ext}$) than were used during training ($T$), the model produces even more compact and separable clusters, significantly boosting performance without additional training overhead.
Experimental Evaluation
The DCG framework was evaluated against eight state-of-the-art baselines across five diverse benchmarks: Synthetic3D, CUB (birds), HandWritten (digits), LandUse-21 (satellite imagery), and Fashion (clothing). The experiments tested missing data rates ranging from 10% to 70%.
Key Results
- Performance at High Missing Rates: DCG demonstrated remarkable robustness in extreme scenarios. On the CUB dataset with 70% of views missing, DCG outperformed the next best method by 6.67% in accuracy.
- Stability: While competing methods often saw drastic performance declines as missingness increased, DCG remained stable. On the HandWritten dataset, DCG's accuracy dropped by only 3% when moving from a 10% to a 70% missing rate, compared to a 9% drop for the top competitor, ICMVC.
- Ablation Studies: Results confirmed that both the diffusion module and the contrastive loss are critical; removing either component led to a significant decrease in clustering performance.
Conclusion
Diffusion Contrastive Generation (DCG) represents a significant advancement in Incomplete Multi-view Clustering. By treating the recovery of missing views as a diffusion-based generative task guided by contrastive objectives, the framework effectively handles high missingness rates while simultaneously improving cluster cohesion. Future work may explore the computational efficiency of the iterative diffusion process and the model's scalability to datasets with a much larger number of views.
Summary updated