Semi-Supervised Nonnegative Matrix Factorization via Constraint Propagation.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Nonnegative matrix factorization (NMF) is a key dimensionality reduction technique.
• Traditional NMF is unsupervised, limiting its use with prior or supervised information.
• Existing semi-supervised NMF methods show limited gains with scarce supervised data.

Purpose of the Study:

• To propose a novel semi-supervised NMF method (CPSNMF) that effectively utilizes limited pairwise constraints.
• To enhance NMF by incorporating must-link and cannot-link information from constrained to unconstrained samples.
• To improve the utilization of constraint information for preserving data distribution geometry.

Main Methods:

• Developed a novel semi-supervised NMF method (CPSNMF) incorporating pairwise constraints.
• Propagated must-link and cannot-link constraints across the entire dataset.
• Integrated constraint information into the data weight matrix and NMF objective function as a regularization term.
• Explored two formulations with corresponding update rules for optimization.

Main Results:

• The proposed CPSNMF method effectively utilizes limited supervised information.
• The method successfully preserves the geometry of the data distribution.
• Experiments demonstrate superior performance compared to existing methods on standard databases.

Conclusions:

• CPSNMF offers a significant advancement in semi-supervised NMF by effectively using pairwise constraints.
• The method enhances dimensionality reduction, particularly when supervised information is limited.
• CPSNMF shows robust performance and broad applicability in areas like computer vision and document clustering.