Multi-graph clustering via multi-modal topological manifold learning.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multi-view spectral clustering is valuable for uncovering hidden data structures.
• Existing methods often use Euclidean distance, limiting their ability to capture relationships between dissimilar-looking but high-similarity samples.
• Early fusion in similarity graph construction leads to information loss.

Purpose of the Study:

• To develop a novel multi-view clustering method, MC_MTML, that overcomes the limitations of existing approaches.
• To accurately measure sample relationships, even for distant but similar data points.
• To reduce information loss during the integration of multi-view data.

Main Methods:

• Generated initial affinity graphs using the K-Nearest Neighbor (KNN) algorithm.
• Learned similarity matrices by incorporating topological manifold learning.
• Derived a common consensus representation for spectral embedding from multiple similarity matrices.
• Developed an efficient alternating iterative algorithm to solve the optimization problem.

Main Results:

• The proposed MC_MTML algorithm effectively incorporates consensus structure knowledge from all feature views.
• The method significantly reduces information loss compared to traditional early fusion techniques.
• Experimental results on toy and real-world multi-view datasets demonstrate superior performance.

Conclusions:

• MC_MTML offers an effective solution for multi-view spectral clustering by addressing limitations in similarity graph construction and data fusion.
• The integration of topological manifold learning enhances the ability to capture complex data relationships.
• The proposed method outperforms existing state-of-the-art multi-view clustering techniques.