代价敏感分类算法的实验比较

An Empirical Comparative Study of Cost-Sensitive Classification Algorithms

对8种不同代价敏感分类算法进行了比较研究。目的是通过实验手段,分析不同代价敏感算法的行为和当其归纳过程发生变化时,对错误分类的总代价、高代价错误数量和错误的总数量所产生的影响。对其中的Ada-Cost方法,本文分析了为何其代价调整因子可能对其性能带来负面影响,并实现了2种变体方法,提高了其性能。

This paper describes a study of different cost-sensitive classification algorithms. The purpose of the study is to analyze the behavior of various cost-sensitive algorithms and how the variations in the induction process affect the total misclassification cost, high cost error amount and total misclassification error amount. For the AdaCost method, this paper analyzes why the cost adjustment factor may cause negative effect on its performance, and implements two modification methods that improve performance substantially.