Semi-Supervised Non-Negative Matrix Factorization With Dissimilarity and Similarity Regularization.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Non-negative matrix factorization (NMF) is a widely used dimensionality reduction technique.
• Traditional NMF often struggles with incorporating label information for supervised tasks.
• Improving clustering performance with limited labeled data remains a challenge.

Purpose of the Study:

• To propose a semi-supervised non-negative matrix factorization (NMF) model that effectively integrates label information.
• To enhance the discriminative power of low-dimensional representations for improved clustering.
• To develop a robust model that outperforms existing NMF-based clustering methods.

Main Methods:

• Developed a semi-supervised NMF model incorporating similarity and dissimilarity regularizers.
• Formulated the model as a constrained optimization problem.
• Solved the problem using an efficient alternating iterative algorithm with theoretical convergence guarantees.

Main Results:

• The proposed NMF model significantly improved clustering accuracy across five benchmark datasets.
• Clustering accuracy increased from 57.0% to 82.2% compared to state-of-the-art methods.
• The model effectively generates discriminative low-dimensional representations by leveraging label information.

Conclusions:

• The proposed semi-supervised NMF model offers a superior approach for clustering tasks.
• The integration of label information via complementary regularizers is highly effective.
• The method provides a significant advancement over conventional NMF techniques in terms of performance.