Abstract:In practice, nonnegative data factorization is often performed for data dimensionality reduction prior to a classification task using a classifier which is effective in low dimensional space such as nearest neighbor classifier. In this work, we propose a novel formulation to learn a multi-class classifier directly through a supervised nonnegative data factorization. This new formulation has the following properties: 1) the nonnegative data matrix is approximated as the product of a nonnegative basis matrix and a coefficient matrix where the nonnegative bases distinctively capture the common characteristics of all classes apart from that specific to individual classes; 2) a regularization term is imposed on nonnegative data factorization so that each datum can be predominantly reconstructed by the common basis vectors and its corresponding class-specific basis vectors; and 3) the coefficient vector for each datum is assumed to be transformed from a mapped kernel space, and the l2 norm of the class-specific coefficients reveals the relative confidence of classes, which then directly leads to a multi-class classifier. We also present an iterative optimization technique for our formulation and analytically show its convergence property. Extensive experiments on face recognition, head pose estimation, and handwritten digit recognition tasks clearly demonstrate the advantages of the proposed classifier over the conventional two-step approach of nonnegative data factorization followed by a classic classifier.

Factorization towards a classifier

Discriminative Concept Factorization for Data Representation

Unified Solution to Nonnegative Data Factorization Problems

Non-Negative Matrix Factorization Framework For Face Recognition

Nonnegative Discriminant Matrix Factorization

Robust discriminative non-negative matrix factorization.

Method of Class-information-incorporated Non-negative Matrix Factorization and Its Application

Large Margin Based Nonnegative Matrix Factorization and Partial Least Squares Regression for Face Recognition

LDA-based Non-negative Matrix Factorization for Supervised Face Recognition

Non-Negative Matrix Factorization and Support Vector Data Description Based One Class Classification

Robust Structured Convex Nonnegative Matrix Factorization for Data Representation

A Discriminant Nonnegative Tensor Factorization Method Based on Sparse Representation Classifier

Class-Driven Non-Negative Matrix Factorization for Image Representation

Discriminative Multiview Nonnegative Matrix Factorization With Large Margin For Image Classification

Nonnegative Tensor Cofactorization and Its Unified Solution

A modified non-negative Matrix Factorization algorithm for face recognition

Robust Semisupervised Nonnegative Local Coordinate Factorization for Data Representation

Concept Factorization by Joint Locality-constrained and l(2,1)-norm Regularization for Image Representation

Discriminant sparse nonnegative matrix factorization

Non-negative Matrix Factorization Via Discriminative Label Embedding for Pattern Classification.

Factor-Bounded Nonnegative Matrix Factorization