Multi-class cost sensitive AdaBoost algorithm based on cost sensitive exponential loss function
Zhai Xiyang, Wang Xiao-dan, Yang Jiang, Wen Tong · 2017
The AdaBoost algorithm which is an important ensemble learning algorithm can effectively improve the classification performance of weak classifiers. Meanwhile the cost sensitive AdaBoost algorithm is an important cost sensitive ME algorithm which can resolve cost sensitive problem effectively. Because the most existing cost sensitive AdaBoost algorithms are binary, a multi-class cost sensitive AdaBoost algorithm based on constructing base classifiers was proposed. The algorithm is complex and the capability relies on the base classifiers seriously. To solve these problems, this paper proposes a multi-class cost sensitive AdaBoost algorithm based on cost exponential loss function. This paper designs a cost sensitive multi-class exponential loss function, and it is proved that the decision function with minimum loss function converges to cost sensitive Bayesian decision function. On this basis, employ the stagewish additive modeling to deduce CSSAMME - a multi-class cost sensitive AdaBoost algorithm. Finally, use UCI dataset to verify the CSSAMME algorithm. The experiment results show that the algorithm has cost sensitive characteristic and the convergence characteristic.