An efficient ADMM-type algorithm for deep semi-nonnegative matrix factorization

Abstract

Abstract In this paper, we focus on deep semi-nonnegative matrix factorization (DSemiNMF) which has a wider application in the real world than traditional NMF. We propose an efficient algorithm based on the classic alternating direction method of multipliers (ADMM) for DSemiNMF. By utilizing structures in DSemiNMF, we derive an efficient updating rule for updating subproblems according to its KKT conditions. Numerical experiments are conducted to compare the proposed algorithm with state-of-the-art deep semi-NMF algorithm. Results show that our algorithm performs better and the deep model indeed results in better clustering accuracy than single-layer model.