一种改进的代价敏感随机森林算法

随机森林在分类不平衡数据时,容易偏向多数类而忽略少数类,可以将代价敏感用于分类器的训练,但在传统代价敏感随机森林算法中,代价函数没有考虑样本集实际分布与特征权重,且在随机森林投票阶段,没有考虑基分类器的性能差异。本文提出一种改进的代价敏感随机森林算法ICSRF,该算法首先根据不平衡数据集的实际分布构造代价函数,并将权重距离引入代价函数,然后根据基分类器的性能采取权重投票,提高分类准确率。实验结果表明,ICSRF算法能有效提高少数类的分类性能,可以较好的处理不平衡数据。

An Improved Cost-sensitive Algorithm Based on Random Forest

The random forest prefers to majority classes rather than minority classes on imbalanced data . The cost sensitive method can be combined with random forest to solve the imbalanced problem . But the traditional cost-sensitive algorithm based on random forest does not consider the actual distribution of data set and feature weight. And in the voting stages of random forest, it does not consider the performance differences of base classifiers. This paper proposed an improved cost-sensitive algorithm based on random forest ICSRF, which constructs a cost function based on the actual distribution of imbalanced data set and introduced the weight distance, then takes weighted voting according to the performance of the base classifier . It can improve the classification accuracy. The experiment results show that the ICSRF algorithm has higher accuracy rate and can effectively improve the recognition rate of the minority classes.