基于加权聚类质心的SVM不平衡分类方法

Support vector machine imbalanced data classification based on weighted clustering centroid

不平衡数据分类是机器学习研究的热点问题,传统分类算法假定不同类别具有平衡分布或误分代价相同,难以得到理想的分类结果.提出一种基于加权聚类质心的SVM分类方法,在正负类样本上分别进行聚类,对每个聚类,用聚类质心和权重因子代表聚类内样本分布和数量,相等类别数量的质心和权重因子参与SVM模型训练.实验结果表明,该方法使模型的训练样本具有较高的代表性,分类性能与其他采样方法相比得到了提升.

Classification of imbalanced data has become a research hot topic in machine learning. Traditional classi- fication algorithms assume that different classes have balanced distribution or equal misclassification cost, thus, making it hard to get ideal result of classifications. A support vector machine （ SVM） classification method based on weighted clustering centroid was proposed in this paper. First, unsupervised clustering was applied to the positive and negative samples respectively to extract the clustering centroid of each clustering, which was represented the most in compactness of the clustering sample. Next, all clustering centroids formed a new set of balance training. In order to minimize the information loss during clustering, each clustering centroid was associated with a weight factor that was defined proportional to the number of samples of the class. Finally, all clustering centroids and weight fac- tors participated in the training of the improved SVM model. Experimental results show that the proposed method can make the sample selected from model train sets more typical and improve the classification performance better than other sampling techniques for dealing with imbalanced data.