TL;DR
This paper introduces STCMC-UR, a novel multi-view clustering method that leverages uncertainty estimation to refine self-supervised contrastive learning. By integrating Dempster-Shafer theory for robust evidence fusion and selectively using pseudo-labels based on confidence, STCMC-UR achieves more reliable and accurate clustering, particularly in critical applications.
What Problem Does This Paper Address?
Multi-view clustering (MVC) aims to group data points by leveraging information from multiple sources (views). Contrastive learning has emerged as a powerful technique for MVC, learning representations by pulling similar samples together and pushing dissimilar ones apart. However, existing contrastive MVC methods often overlook the reliability of their clustering results and are susceptible to "false negatives" – incorrectly treating dissimilar samples as similar. This lack of robustness limits their application in security-sensitive domains. Furthermore, these methods typically treat all views equally, ignoring their varying importance and potential for introducing noise.
Methods
STCMC-UR tackles these challenges through a multi-module approach:
- Evidence Generation Module: This module generates "evidence" for each view using two distinct Basic Belief Assignment (BBA) methods:
- BBA-SP: Uses the Softplus function for a more reliable output than softmax, avoiding over-confidence.
-
BBA-SM: Reallocates deep neural network outputs using softmax, providing greater certainty for some predictions. These two BBAs are fused using the Dirichlet distribution and Dempster-Shafer theory to produce a more robust view-specific belief mass and uncertainty.
-
Multi-view Fusion: The Dempster-Shafer theory is employed again to fuse the evidence from all views into a single, unified clustering result and its associated uncertainty. This fusion process is designed to be robust, giving more weight to views with higher confidence.
- View Weighting: To account for the varying importance of different views, the method learns view weights based on the clustering results. These weights are then used to adjust the evidence, further emphasizing reliable information and downplaying noisy contributions.
- Uncertainty-Refined Self-supervised Contrastive Learning: This is the core innovation. Instead of using all generated pseudo-labels for contrastive learning, STCMC-UR uses the estimated uncertainty. A threshold ($\epsilon$) is set: if a sample's uncertainty is below this threshold (indicating high confidence), its pseudo-label is used for contrastive learning. Otherwise, it's discarded. This selective pseudo-labeling helps to filter out false negatives, leading to more accurate feature alignment and improved clustering.
Datasets
The proposed STCMC-UR method was evaluated on five publicly available multi-view datasets:
- Caltech-2V
- Caltech-3V
- COIL20
- Event8
- NUS22
These datasets vary in the number of views, samples, clusters, and feature dimensionality, providing a comprehensive evaluation.
Key Insights and Ideas
- Uncertainty as a Guide: The paper highlights that simply generating clustering results isn't enough; understanding their reliability (uncertainty) is crucial for effective learning.
- Dempster-Shafer for Robust Fusion: Using Dempster-Shafer theory for both within-view and across-view fusion provides a principled way to combine evidence from multiple sources, leading to more stable and trustworthy results.
- Selective Pseudo-labeling for CL: The most significant insight is using uncertainty to filter pseudo-labels for contrastive learning. This directly addresses the problem of false negatives, which can significantly degrade performance in traditional contrastive methods.
- Mutual Benefit of Modules: The evidence generation, fusion, and contrastive learning modules are designed to be mutually beneficial, with each module's output informing and improving the others.
Key Results
STCMC-UR demonstrated significant improvements over state-of-the-art methods across all five datasets.
- On Caltech-2V, it achieved a 7.3% increase in ACC and a 15.5% increase in NMI compared to the second-best method.
- On the challenging COIL20 dataset, STCMC-UR achieved perfect clustering results (100% ACC and NMI).
- Ablation studies confirmed that each component (evidence generation, view fusion, contrastive learning) contributes to the overall performance gain.
- Parameter sensitivity analysis showed that the method is robust to variations in its hyper-parameters ($\alpha$ and $\beta$).
- Convergence analysis indicated that the model converges efficiently within approximately 40-50 epochs.
Limitations and Future Work
The authors acknowledge a key limitation: the proposed method fails to deal with incomplete multi-view data, where missing values in certain views create imbalances in information content across views. They plan to address this limitation in future work.
Summary updated