A Comparison of Weighted Support Vector Machine (WSVM), One-Step WSVM (OWSVM) and Iteratively WSVM (IWSVM) for Mislabeled Data

摘要

Labeling error can occur for various reasons such as the subjective nature of the labeling task, the lack of information to determine the true label of a given example and data entry error. Labeling errors were categorized as mislabeled, unlabeled, partially labeled, incompletely labeled and illegible label. In this study, the focus will be on mislabeled data. The problem of dealing with mislabeled data and in particular of constructing a classifier from such data has been approached from a number of different directions. Therefore, developing learning algorithms that effectively and efficiently deal with mislabeled data is a great practical importance and key aspect in machine learning. Support Vector Machine (SVM) has been widely accepted to be one of the most effective techniques in machine learning algorithms. One of the main drawbacks of SVM is it depends on only a small part of the data points (support vectors) and it treats all training data of a given class equally. To address this problem, one of the solution is the Weighted Support Vector Machines (WSVM). Wu & Liu proposed two different WSVM namely one-step WSVM (OWSVM) and iteratively WSVM (IWSVM). In this paper, a comparison of Weighted Support Vector Machine (WSVM), One-step WSVM (OWSVM) and Iteratively WSVM (IWSVM) for mislabeled data has been done to see the classification accuracy of each of the method. The three methods were compared based on correctly labeled, mislabeled data, data within margin, mislabeled data within margin and classification accuracy for eight KEEL repository datasets using 20% noise in training data. Based on the experimental results, the performance of OWSVM is better than both WSVM and IWSVM based on the correctly labeled, mislabeled data, data within margin, mislabeled data within margin and classification accuracy.