Convmd: Convolutive Matrix Decomposition For Classification Of Matrix Data

Abstract

In this paper, we consider the use of convolutive matrix decomposition for matrix data classification. Matrix decomposition has been broadly used as means of dimensionality reduction in a variety of learning tasks. In this approach, columns of a matrix are represented as a linear combination over a basis. For applications in which relevant information is encoded in a sequence of columns instead of a single column, the use of a single column basis is insufficient. In this paper, we present a matrix classification framework that relies on a convolutive-based matrix decomposition approach that captures structure among neighboring columns. In particular, we present a latent variable graphical model for classification of matrices that is based on the proposed matrix decomposition. We present experimental results with promising performance on a DNA dataset associated with protein production.