A New Loss Function For Robust Classification

Lei Zhao; Musa Mammadov; John Yearwood

doi:10.3233/IDA-140664

Back

A New Loss Function For Robust Classification

Journal article

Open access

Peer reviewed

A New Loss Function For Robust Classification

Lei Zhao, Musa Mammadov and John Yearwood

Intelligent Data Analysis, Vol.18, pp.697-715

2014

DOI: https://doi.org/10.3233/IDA-140664

Files and links (2)

pdf

PDF - Author's Accepted Version308.08 kBDownload View

Accepted VersionPDF - Author Accepted Version Open Access

url

https://doi.org/10.3233/IDA-140664View

Published Version

Abstract

classification

loss function

machine learning

optimization

data mining

Loss function plays an important role in data classification. Many loss functions have been proposed and applied to different classification problems. This paper proposes a new so called the smoothed 0-1 loss function, that could be considered as an approximation of the classical 0-1 loss function. Due to the non-convexity property of the proposed loss function, global optimization methods are required to solve the corresponding optimization problems. Together with the proposed loss function, we compare the performance of several existing loss functions in the classification of noisy data sets. In this comparison, different optimization problems are considered in regards to the convexity and smoothness of different loss functions. The experimental results show that the proposed smoothed 0-1 loss function works better on data sets with noisy labels, noisy features, and outliers.

Details

Title: A New Loss Function For Robust Classification
Authors: Lei Zhao (Author) - University of the Sunshine Coast
Musa Mammadov (Author) - University of Ballarat
John Yearwood (Author) - University of Ballarat
Publication details: Intelligent Data Analysis, Vol.18, pp.697-715
Publisher: IOS Press
Date published: 2014
DOI: 10.3233/IDA-140664
ISSN: 1088-467X; 1088-467X
Organisation Unit: Insights & Analytics Unit; University of the Sunshine Coast, Queensland; Office of Research
Language: English
Record Identifier: 99448847102621
Output Type: Journal article
Research Statement: false

Metrics

218 File views/ downloads

1807 Record Views

InCites Highlights

These are selected metrics from InCites Benchmarking & Analytics tool, related to this output

Collaboration types: Industry collaboration; Domestic collaboration
Web Of Science research areas: Computer Science, Artificial Intelligence