Adversarial Graph Regularized Deep Nonnegative Matrix Factorization for Data Representation

Abstract

This work proposes a novel unsupervised deep non-negative matrix factorization (NMF) model called AGDNMF by deep exploration of the structure of the original data. Compared with the existing NMF research results, the model explores the deep association of data and constructs accurate spatial structure from a new perspective. Specifically, for the purpose of effectively learning low-dimensional representations of data. This work constructs a deep NMF method to approximate the coding structure of the original data, which enables AGDNMF to learn the hierarchical mapping relationship between the categories of the original data, and learn the hidden association information of the data from low-to-high in the middle layer. Moreover, the model gives an adversarial graph regularization representation for the manifoldization of data structures to push away the spatial distances between different sample clusters. Finally, based on the mathematical model of the proposed algorithm, the optimization strategy and convergence proof are given. Extensive experiments and data analysis demonstrate that the proposed model outperforms state-of-the-art unsupervised NMF methods.

Keywords

• Deep nonnegative matrix factorization
• adversarial graph regularization
• unsupervised learning
• data representation