Using Non-negative Matrix Factorization for Bankruptcy Analysis

Abstract

Dimensionality reduction is demonstrated crucial to improve the predictive capability of models by means of linear or nonlinear projections. Non-negative matrix factorization (NMF) is a popular multivariate analysis technique for part-based data representation. It attempts to find an approximation of a high dimensional matrix as the product of two low dimensional matrices under the non-negative constraint. Recently a graph regularized non-negative matrix factorization (GNMF) provides a formal way to incorporate the geometrical structure into the NMF decomposition, particularly applicable to the data embedded in submanifolds of the Euclidean space. In this paper, the usage of GNMF in financial analysis is discussed from the perspectives of unsupervised clustering and supervised classification. Experimental results on a French bankruptcy data set show the potential of GNMF on data representation.