Sparse Nonnegative Matrix Factorization Based on Collaborative Neurodynamic Optimization

This paper presents a collaborative neurodynamic approach to sparse nonnegative matrix factorization (SNMF). SNMF is formulated as a bilevel optimization problem. In the lower level of the problem, the sparsity of factorized matrix is minimized subject to the factorization error and nonnegative constraints. In the upper level of the problem, the parameter of the inverted Gaussian function is minimized to approximate l
₀
norm. A group of neurodynamic models operating at two timescales is employed to solve the reformulated problem. The experimental results show the superiority of the proposed approach.