Selective Cross-View Topology for Deep Incomplete Multi-View Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Incomplete multi-view data is common in real-world applications.
• Existing methods often fail to utilize inter-view relationships effectively.
• Unsupervised learning settings require robust methods for handling missing data across views.

Purpose of the Study:

• To propose a novel framework for incomplete multi-view clustering that addresses the limitations of existing methods.
• To effectively leverage cross-view topological relationships for view completion and representation learning.
• To improve clustering performance on datasets with missing multi-view information.

Main Methods:

• Constructing a view topology graph using Optimal Transport (OT) distance to identify neighboring views.
• Implementing a Max View Graph Contrastive Alignment module for information transfer across views.
• Utilizing a View Graph Weighted Intra-View Contrastive Learning module for enhanced representation learning.

Main Results:

• The proposed Selective Cross-View Topology Incomplete Multi-View Clustering (SCVT) framework achieves state-of-the-art performance.
• SCVT significantly outperforms existing methods on seven benchmark datasets.
• The method demonstrates effectiveness in both view completion and representation learning for incomplete multi-view data.

Conclusions:

• Leveraging selective cross-view topological relationships is crucial for effective incomplete multi-view clustering.
• The SCVT framework provides a robust solution for handling missing multi-view data.
• The proposed approach enhances clustering accuracy and representation learning through graph-based alignment and contrastive learning.