One Class Support Vector Machine Active Learning Method for Unbalanced Data

Abstract

Active learning is an important method to efficiently reduce the labeled sample size in large-scale machine learning. As the large amount of data presents the nature of imbalance, the study of active learning methods for corresponding unbalanced datasets highlights important practical significance. There is a problem with the current algorithm. The initialization of active learning becomes difficult with the increase of the unbalanced degree, which causes more labels waste. This paper proposes a one-class support vector machine active learning (OCSVM-AL) method for this problem. Firstly, we obtain the structure information of the sample set by k-means clustering. Then, the OCSVM-AL is used on each cluster. For each cluster, OCSVM-AL iteratively marks the sample from the cluster and adds the marked sample to the marked sample pool. A one-class support vector machine model is trained by using majority class samples in the marked sample pool. The trained model is used to predict the unlabeled samples in this cluster, and the labels of unlabeled samples is obtained. By the results of each cluster, all sample labels are obtained. Simulation experiments proved that this method is an effective active learning method and can significantly reduce the labeled sample size.