Bicriteria Sparse Nonnegative Matrix Factorization via Two-Timescale Duplex Neurodynamic Optimization.

Area of Science:

• * Computational Mathematics
• * Machine Learning
• * Optimization

Background:

• * Sparse Nonnegative Matrix Factorization (SNMF) is crucial for data analysis, but optimizing for both low error and high sparsity is challenging.
• * Existing methods often struggle to balance factorization accuracy with the degree of sparsity achieved.

Purpose of the Study:

• * To formulate SNMF as a bicriteria optimization problem using an exact binary representation of the l0 matrix norm.
• * To convert the problem into a solvable biconvex form.
• * To introduce a novel two-timescale duplex neurodynamic approach for solving the biconvex SNMF problem.

Main Methods:

• * Formulation of SNMF as a mixed-integer bicriteria optimization problem.
• * Transformation of binary constraints into equivalent bilinear constraints, creating a biconvex problem.
• * Implementation of a two-timescale duplex neurodynamic system with two collaborative Recurrent Neural Networks (RNNs).
• * Definition of a Gaussian Score (GS) to integrate factorization error and sparsity.

Main Results:

• * The proposed neurodynamic approach successfully solved the biconvex SNMF problem.
• * Demonstrated low factorization errors and high matrix sparsity across four benchmark datasets.
• * Achieved high Gaussian Scores, indicating a superior balance between error minimization and sparsity maximization.

Conclusions:

• * The developed two-timescale duplex neurodynamic approach offers an effective solution for sparse nonnegative matrix factorization.
• * The method provides a robust way to simultaneously minimize errors and maximize sparsity.
• * The approach shows significant promise for applications requiring highly sparse and accurate matrix factorizations.