Manifold Based Multi-View K-Means.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• K-means clustering is widely used but struggles with accurate center estimation and linearly non-separable data.
• Existing multi-view K-means methods often depend on centroid calculations, posing optimization challenges.

Purpose of the Study:

• To develop a novel multi-view K-means clustering model that addresses limitations of traditional K-means.
• To enhance clustering performance by incorporating manifold learning and tensor rank constraints.

Main Methods:

• Reformulated multi-view K-means from a manifold learning perspective, eliminating the need for centroid matrices.
• Proposed a new model using indicator matrices from different views to construct a third-order tensor.
• Applied tensor Schatten p-norm to minimize tensor rank, effectively utilizing complementary information across views.
• Incorporated diverse distance functions to handle linearly non-separable data.

Main Results:

• The proposed model ensures consistency between manifold structure and data labels.
• Demonstrated superior performance compared to existing methods on multiple benchmark datasets.
• Effectively leveraged complementary information across multiple data views.

Conclusions:

• The novel manifold-based multi-view K-means model with tensor rank constraint offers a robust solution for clustering challenges.
• The method shows significant improvements in handling complex, non-separable data structures.
• This approach advances multi-view clustering by integrating manifold learning and tensor analysis.