摘要： 现实世界中广泛存在着很多不均衡的数据,其分类问题是机器学习领域的研究热点。为了提高不均衡数据的分类性能,提出一种基于核空间置信度的代价敏感支持向量机分类算法。通过注入类别错分代价机制,以不均衡数据评价指标作为目标函数,优化错分代价因子,提升少数类样本的识别率。计算类中所有样本在核空间下的类别置信度,从而确定样本对决策分类贡献的重要程度,降低噪音或孤立点对支持向量机的影响。通过大量UCI数据集的实验结果表明,与其他同类算法相比,该算法能更好地提高不均衡数据的分类性能。

关键词: 机器学习, 分类, 不均衡数据学习, 支持向量机, 代价敏感学习

Abstract: Imbalanced data classification problem is one of the main research field of machine learning in the real world.In order to improve the classification performance of Support Vector Machine(SVM),a kernel space confidence based cost SVM is proposed.It can improve the accuracy of minority class by injecting the strategy of misclassification cost into training.Using the imbalanced data evaluation metric as the objective function,the method optimizes the misclassification cost parameter,so as to improve the accuracy of minority class.Moreover,the weight of each instance for decision classification contribution can be obtained by calculating the class confidence on the kernel space,so as to decrease the effect of noisy and outlier instances for SVM.Experimental results show that the proposed algorithm provides a very competitive solution to other existing methods for combating imbalanced classification problems.

Key words: machine learning, classification, imbalanced data learning, Support Vector Machine(SVM), cost sensitive learning

中图分类号: 

• TP18