A deep structure-enforced nonnegative matrix factorization for data representation

Abstract

In this paper, we focus on a deep structure-enforced nonnegative matrix factorization (DSeNMF) which represents a large class of deep learning models appearing in many applications. We present a unified algorithm framework, based on the classic alternating direction method of multipliers (ADMM). For updating subproblems, we derive an efficient updating rule according to its KKT conditions. We conduct numerical experiments to compare the proposed algorithm with state-of-the-art deep semi-NMF. Results show that our algorithm performs better and our deep model with different sparsity imposed indeed results in better clustering accuracy than single-layer model. Our DSeNMF can be flexibly applicable for data representation.