Partial Contrastive Learning for Partially View-aligned Multi-view Clustering

Area of Science:

• Machine Learning
• Data Science

Background:

• Deep neural networks have advanced multi-view clustering.
• Existing methods often assume complete cross-view correspondence, which is unrealistic.

Purpose of the Study:

• To address partially aligned multi-view clustering.
• To propose a novel framework, Partial Contrastive Learning (PCL).

Main Methods:

• PCL combines explicit cross-view correspondence modeling with contrastive learning.
• A partial alignment module adaptively computes soft matching probabilities.
• These probabilities guide a generalized contrastive loss for robust representation learning.

Main Results:

• PCL enhances the discriminative power of multi-view representations.
• Achieved superior alignment and clustering performance.
• Demonstrated effectiveness across eight benchmark datasets against eleven methods.

Conclusions:

• PCL effectively handles partially aligned multi-view data.
• The framework improves both data alignment and clustering accuracy.
• PCL offers a practical solution for real-world multi-view clustering challenges.