Symmetric nonnegative matrix factorization: algorithms and applications to probabilistic clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computational Biology

Background:

• Nonnegative matrix factorization (NMF) is a powerful unsupervised learning technique.
• Symmetric NMF (SNMF) is a specialized variant with broad applications.
• Existing NMF methods can be computationally intensive.

Purpose of the Study:

• To develop efficient parallel algorithms for SNMF.
• To improve the performance of SNMF in clustering tasks.
• To demonstrate the applicability of SNMF algorithms in diverse domains.

Main Methods:

• Development of three parallel multiplicative update algorithms for SNMF.
• Algorithms utilize level 3 basic linear algebra subprograms for efficiency.
• Convergence analysis of the proposed Euclidean distance minimization algorithm.

Main Results:

• Successful implementation of three novel parallel SNMF algorithms: α-SNMF and β-SNMF.
• Demonstrated effectiveness in probabilistic clustering tasks.
• Validated performance in facial image clustering, document categorization, and gene expression pattern clustering.

Conclusions:

• The proposed parallel SNMF algorithms are effective and easy to implement.
• These algorithms offer significant improvements for various data clustering applications.
• SNMF provides a robust framework for unsupervised learning and pattern recognition.