Semi-Supervised Structured Subspace Learning for Multi-View Clustering.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multi-view clustering seeks a consensus subspace for clustering across diverse data sources.
• Traditional methods often lack semi-supervised capabilities and robust handling of inter-view relationships.

Purpose of the Study:

• To propose a novel semi-supervised multi-view clustering algorithm, Semi-supervised Structured Subspace Learning for Multiple sources (SSSL-M).
• To enhance clustering by learning a comprehensive shared affinity matrix that integrates complementary information from multiple views.

Main Methods:

• Developed SSSL-M, integrating semi-supervised learning with structured subspace learning.
• Employed an anti-block-diagonal indicator matrix to enforce block-diagonal structure in the shared affinity matrix.
• Utilized backward encoding networks and self-expressive mapping within a unified framework for affinity matrix regularization.

Main Results:

• SSSL-M demonstrated superior clustering performance across seven benchmark datasets compared to state-of-the-art methods.
• The proposed method effectively captures complementary information and enhances structural consistency among affinity matrices from different views.
• Experimental validation confirmed the efficacy of the alternating optimization scheme for solving the formulated problem.

Conclusions:

• SSSL-M offers a powerful semi-supervised approach for multi-view clustering, outperforming existing techniques.
• The method's ability to learn a comprehensive shared affinity matrix is key to its improved clustering accuracy.
• The structured subspace learning framework effectively leverages supervisory information for enhanced multi-view data analysis.