Manifold Peaks Nonnegative Matrix Factorization.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Nonnegative Matrix Factorization (NMF) is increasingly recognized for its interpretability in data analysis.
• NMF is closely linked to fuzzy k-means clustering, with basis matrices representing cluster centroids.
• Existing NMF clustering methods often yield centroids that deviate from the data manifold, compromising results on complex datasets.

Purpose of the Study:

• To introduce Manifold Peaks Nonnegative Matrix Factorization (MPNMF) for enhanced data clustering.
• To address the limitation of centroid deviation from the data manifold in traditional NMF clustering.
• To improve clustering accuracy on datasets with intricate geometric structures.

Main Methods:

• MPNMF selects Manifold Peaks (MPs) to define the data manifold's structure.
• Centroids are constrained to lie on the data manifold via conic combinations of nearby MPs.
• Graph smoothness regularization is generalized for local graph construction.
• The method solves a quadratic regularized nonnegative least squares problem with a group l0-norm constraint, utilizing an efficient optimization algorithm.

Main Results:

• MPNMF effectively constrains centroids to the data manifold.
• The approach demonstrates superior performance compared to existing methods on synthetic and real-world datasets.
• The generalized graph smoothness regularization aids in constructing informative local graphs.

Conclusions:

• MPNMF offers a robust solution for data clustering, particularly for datasets with complex geometric properties.
• By keeping centroids on the data manifold, MPNMF enhances the reliability and interpretability of clustering results.
• The proposed optimization algorithm efficiently solves the MPNMF objective function, making it practical for various applications.