Semisupervised Adaptive Symmetric Non-Negative Matrix Factorization.

Area of Science:

• Machine Learning
• Data Mining
• Computational Science

Background:

• Symmetric Non-negative Matrix Factorization (SymNMF) offers direct clustering but relies on predefined similarity matrices, limiting performance.
• The quality of the similarity graph is critical for effective clustering outcomes in SymNMF.

Purpose of the Study:

• To develop a novel semisupervised model that jointly learns the similarity matrix and generates clustering results.
• To leverage pairwise constraints to propagate supervisory information for a more informative similarity matrix.
• To enhance clustering performance through the synergistic effect of learning the similarity matrix and clustering simultaneously.

Main Methods:

• A semisupervised model is proposed to simultaneously learn the similarity matrix and perform clustering.
• The model utilizes pairwise constraints to propagate supervisory information, creating an informative similarity matrix.
• The problem is formulated as a non-negativity-constrained optimization problem, solved by a proposed iterative method with proven convergence.

Main Results:

• The proposed semisupervised model demonstrates superior clustering performance compared to traditional methods.
• Extensive experiments validate the effectiveness of the model in learning informative similarity matrices.
• The iterative solution method ensures convergence and reliable performance.

Conclusions:

• The proposed semisupervised approach significantly improves clustering performance by integrating similarity learning with clustering.
• The model effectively utilizes supervisory information to enhance the similarity matrix, leading to better clustering results.
• This method offers a robust and theoretically sound approach to semisupervised clustering using NMF variants.