White Rose University Consortium logo
University of Leeds logo University of Sheffield logo York University logo

A sparse multinomial probit model for classification

Ding, Y. and Harrison, R.F. (2011) A sparse multinomial probit model for classification. Pattern Analysis and Applications , 146 (1). pp. 47-55. ISSN 1433-7541

Full text available as:
[img]
Preview
Text
Harrison_42898.pdf

Download (219Kb)

Abstract

A recent development in penalized probit modelling using a hierarchical Bayesian approach has led to a sparse binomial (two-class) probit classifier that can be trained via an EM algorithm. A key advantage of the formulation is that no tuning of hyperparameters relating to the penalty is needed thus simplifying the model selection process. The resulting model demonstrates excellent classification performance and a high degree of sparsity when used as a kernel machine. It is, however, restricted to the binary classification problem and can only be used in the multinomial situation via a one-against-all or one-against-many strategy. To overcome this, we apply the idea to the multinomial probit model. This leads to a direct multi-classification approach and is shown to give a sparse solution with accuracy and sparsity comparable with the current state-of-the-art. Comparative numerical benchmark examples are used to demonstrate the method.

Item Type: Article
Copyright, Publisher and Additional Information: © 2011 Springer. This is an author produced version of a paper subsequently published in Pattern Analysis and Applications. Uploaded in accordance with the publisher's self-archiving policy.
Keywords: Multi-classification; Sparseness; Multinomial probit; Hierarchical Bayesian
Academic Units: The University of Sheffield > Faculty of Engineering (Sheffield) > Department of Automatic Control and Systems Engineering (Sheffield)
Depositing User: Miss Anthea Tucker
Date Deposited: 07 Mar 2011 16:14
Last Modified: 08 Feb 2013 17:31
Published Version: http://dx.doi.org/10.1007/s10044-010-0177-7
Status: Published
Publisher: Springer
Refereed: Yes
Identification Number: 10.1007/s10044-010-0177-7
URI: http://eprints.whiterose.ac.uk/id/eprint/42898

Actions (login required)

View Item View Item