论文部分内容阅读
设计直接优化不平衡准则算法是克服SVM在不平衡数据集上表现不佳的一个有效途径.但已有研究多面向F1、AUC等常见标准,对于其他标准如G-TP/PR等由于优化上的困难一直鲜有涉及.对此,提出一个直接优化G-TP/PR的新型算法.算法采用传统SVM框架,定义了面向G-TP/PR的目标函数,该目标比已有基于F1目标更加紧凑.针对新函数非光滑,难以直接优化,提出使用束方法进行求解,使得算法的迭代次数不依赖于训练样本数,更适合大规模的应用场合.不平衡数据集上实验证明了所提算法的有效性.
Designing algorithms to directly optimize unbalanced criteria is an effective way to overcome the poor performance of SVMs on unbalanced datasets, but many common standards such as F1 and AUC have been studied. For other standards such as G-TP / PR, , This paper proposes a new algorithm to directly optimize G-TP / PR.The traditional SVM framework is adopted to define the objective function for G-TP / PR, which is more than the existing goal based on F1 Aiming at the problem that the new function is nonsmooth and difficult to be directly optimized, the beam method is proposed to solve the problem, which makes the iteration number of the algorithm not depend on the number of training samples and is more suitable for large-scale applications.Experiments on unbalanced datasets prove that the proposed algorithm Effectiveness.