文献类型：期刊文献

英文题名：Entropy-based matrix learning machine for imbalanced data sets

作者：Zhu, Changming[1];Wang, Zhe[2]

机构：[1]Shanghai Maritime Univ, Coll Informat Engn, Shanghai 201306, Peoples R China;[2]East China Univ Sci & Technol, Coll Informat Sci & Engn, Shanghai 200237, Peoples R China

年份：2017

卷号：88

起止页码：72

外文期刊名：PATTERN RECOGNITION LETTERS

收录：;EI(收录号：20170503309594);WOS:【SCI-EXPANDED(收录号:WOS:000396957800011)】；

基金：This work was supported by Shanghai Natural Science Foundation under grant numbers 16ZR1414500, Natural Science Foundation of China under Grant Nos. 61602296 and 61672227, and the author would like to thank their supports.

语种：英文

外文关键词：Entropy; Fuzzy membership; Imbalanced data set; Pattern recognition

摘要：Imbalance problem occurs when negative class contains many more patterns than that of positive class. Since conventional Support Vector Machine (SVM) and Neural Networks (NN) have been proven not to effectively handle imbalanced data, some improved learning machines including Fuzzy SVM (FSVM) have been proposed. FSVM applies a fuzzy membership to each training pattern such that different patterns can give different contributions to the learning machine. However, how to evaluate fuzzy membership becomes the key point to FSVM. Moreover, these learning machines present disadvantages to process matrix patterns. In order to process matrix patterns and to tackle the imbalance problem, this paper proposes an entropy-based matrix learning machine for imbalanced data sets, adopting the Matrix-pattern oriented Ho-Kashyap learning machine with regularization learning (MatMHKS) as the base classifier. The new leaning machine is named EMatMHKS and its contributions are: (1) proposing a new entropy-based fuzzy membership evaluation approach which enhances the importance of patterns, (2) guaranteeing the importance of positive patterns and get a more flexible decision surface. Experiments on real-world imbalanced data sets validate that EMatMHKS outperforms compared learning machines. (C) 2017 Elsevier B.V. All rights reserved.