A Discrete-Time Neurodynamic Approach to Sparsity-Constrained Nonnegative Matrix Factorization.

Area of Science:

• Computational mathematics
• Machine learning
• Optimization

Background:

• Nonnegative matrix factorization (NMF) often requires sparse solutions for improved interpretability and performance.
• Existing regularization methods for NMF sparsity are sensitive to parameter tuning and can be suboptimal.
• Achieving high sparsity in NMF solutions remains a challenge in various applications.

Purpose of the Study:

• To formulate sparse NMF as a mixed-integer optimization problem with explicit binary constraints for sparsity.
• To develop and analyze a discrete-time projection neural network for solving the formulated sparse NMF problem.
• To demonstrate the effectiveness of the proposed method in extracting highly sparse features.

Main Methods:

• Formulation of sparse NMF as a mixed-integer optimization problem.
• Development of a discrete-time projection neural network for solving the optimization problem.
• Analytical characterization of network stability and convergence using Lyapunov's method.
• Experimental evaluation on sparse feature extraction tasks.

Main Results:

• The proposed mixed-integer optimization formulation effectively enforces sparsity through binary constraints.
• The discrete-time projection neural network demonstrates stability and convergence properties.
• Experimental results show superior performance in extracting highly sparse features compared to existing methods.
• The approach offers a more principled way to control sparsity in NMF.

Conclusions:

• The presented mixed-integer optimization and projection neural network approach provides an effective solution for sparse NMF.
• This method overcomes limitations of traditional regularization by directly incorporating sparsity constraints.
• The approach is validated for its ability to extract highly sparse features, showing significant advantages in practical applications.