Pairwise Constraint Propagation-Induced Symmetric Nonnegative Matrix Factorization.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Symmetric Nonnegative Matrix Factorization (SNMF) effectively captures graph structure for clustering.
• Existing SNMF methods often rely on empirical similarity matrix construction and rigid incorporation of supervisory information.

Purpose of the Study:

• To propose a novel semisupervised clustering method, Pairwise Constraint Propagation-induced SNMF (PCPSNMF).
• To enable adaptive and simultaneous learning of similarity and assignment matrices.
• To flexibly integrate pairwise constraints for guiding similarity matrix construction.

Main Methods:

• Formulation of a single-constrained optimization problem for PCPSNMF.
• Adaptive and simultaneous learning of similarity and assignment matrices.
• Development of an efficient alternating iterative algorithm for optimization with proven convergence.
• Flexible integration of pairwise constraints to guide similarity matrix construction.

Main Results:

• PCPSNMF demonstrates adaptive and simultaneous learning of both similarity and assignment matrices.
• The method achieves mutual refinement of matrices through communication until convergence.
• Experimental results on benchmark image datasets show superior clustering performance.
• PCPSNMF exhibits reduced sensitivity to initialization compared to state-of-the-art methods.

Conclusions:

• PCPSNMF offers an effective semisupervised clustering approach by adaptively learning matrices.
• The proposed method enhances clustering performance and robustness.
• PCPSNMF provides a flexible and efficient framework for semisupervised graph-based clustering.