单分类支持向量机用于样本不平衡数据集建模研究

Support Vector Machine for Modeling of Sample Imbalance Data Sets

主要研究应用单分类支持向量机(One-class Support Vector Machine,OneSVM)在不平衡样本数据集的建模问题。首先从UCI获得Abalon数据集,随机挑选400正样本和400负样本构建训练集,交叉验证方法用于OneSVM分类器训练,模型对数据集的预测精度达到98.00%,与标准SVM分类器对此数据集83.10%的预测正确率相比具有明显的竞争力。然后在样本数据不平衡数据机上训练加权SVM对负样本的预测精度为76.70%,模型对负样本不具有稳定性。实验结果表明单分类支持向量机在样本数目失衡的学习问题中具有良好的泛化能力,同时单分类支持向量机只用一类样本训练分类器,在算法的复杂度上也具有优势。

The purpose of this article is to study the performance of one-class support vector machine(OneSVM)used in unbalanced dataset prediction.Firstly,the Abalon dataset was obtained for UCI,in which 400 positive and negative samples were selected to construct training datasets,cross-validation method was employed to train OneSVM classifiers,98.00%accuracy was obtained and the results indicates more promisingcompared with the accuracy 83.10%obtained by using SVM methods.Secondly,an unbalanced dataset was constructed to train right-SVM;the specificity accuracy 76.70%showed the model was lack of stability.The results indicate OneSVM has great generalization in unbalanced dataset learning problems,where,OneSVM is only trained on single class samples that intend to more superior in the complexity.