不平衡数据学习算法在相似性学习中的应用

摘要

In the real_world problems, there is an imbalance in the paired_samples. The number of the paired_samples in similarity set is much smaller than the number of the paired_samples in dissimilarity set. To solve this problem, two approaches, dissimilar K nearest neighbor and similar K nearest neighbor ( DKNN_SKNN) and dissimilar K nearest neighbor and similar K farthest neighbor ( DKNN_SKFN) , are proposed to construct paired_samples. Thus, the number of paired_samples in similarity learning is effectively decreased, the training process of SVM is accelerated, and the imbalanced data problem is solved to some degree. In the experiments, the proposed approaches are compared with some standard resampling methods. The results show that the proposed approaches have better performance.%在现实问题中，相似性学习的样本对存在不平衡现象，即相似性样本对的数量会远小于不相似性样本对的数量。针对此问题，文中提出两种样本对构造方法———不相似K近邻-相似K近邻( DKNN_SKNN)和不相似K近邻-相似K远邻( DKNN_SKFN)。运用这两种方法可有针对性地选择相似性学习样本对，不仅可加快支持向量机的训练过程，而且在一定程度上解决样本对之间的不平衡问题。在多个数据集上进行文中方法和经典的重采样方法的对比实验，结果表明DKNN_SKNN和DKNN_SKFN具有良好性能。