Performance enhanced Boosted SVM for Imbalanced datasets

doi:10.1016/j.asoc.2019.105601

2019-10-01

PRE

M. Punniyamoorthy *

DOI：10.1016/j.asoc.2019.105601

原文链接

原文求助

摘要

En 中文

In recent decades, researchers have proposed different boosting algorithms to improve the learning abilities of classification techniques for imbalanced data. Especially for the SVM, Wang et al. (2010) proposed a cost-sensitive boosting algorithm, named Wang's Boosted SVM (WBSVM). In this study we made two major modifications to WBSVM algorithm to improve the classification performance without increasing its existing time complexity (O(n(3))): (i) a distance-based weight-update rule (ii) a sign based classifier removal technique to obtain a performance enhanced classifier ensemble. We used 43 imbalanced benchmark datasets to assess the performance. The results revealed that both the modifications made a significant improvement in the boosting performance. Also, we compared our modified Boosted SVM (MBSVM) performance with K-Nearest Neighbour, Decision Tree, Artificial Neural Network and one of the deep learning techniques - Deep Neural Network performances. We also developed a framework to construct Receiver operating characteristic (ROC) curve for Boosted SVM. We compared both the boosted SVMs with Area under the ROC curve (AUC) values and also in terms of running time of algorithms. Furthermore, MBSVM was applied to multi-criteria inventory classification data since it is imbalanced in nature. The inventory classification outcomes of four benchmark techniques were used as class labels. The performance of MBSVM was compared with those of WBSVM algorithm as well as the traditional SVM technique. (C) 2019 Elsevier B.V. All rights reserved.

Keyword:

Boosted support vector machine

Imbalanced data learning

Deep neural network

Receiver operating characteristic curve

Multi-criteria inventory classification