代价敏感Boosting算法研究

Cost-sensitive boosting algorithms

针对代价敏感学习问题,研究boosting算法的代价敏感扩展。提出一种基于代价敏感采样的代价敏感boosting学习方法,通过在原始boosting每轮迭代中引入代价敏感采样,最小化代价敏感损失期望。基于上述学习框架,推导出两种代价敏感boosting算法,同时,揭示并解释已有算法的不稳定本质。在加州大学欧文分校(University of California,Irvine,UCI)数据集和麻省理工学院生物和计算学习中心(Center for Biological&Computational Learning,CBCL)人脸数据集上的实验结果表明,对于代价敏感分类问题,代价敏感采样boosting算法优于原始boosting和已有代价敏感boosting算法。

In terms of the problem of cost-sensitive learning,this paper investigates cost-sensitive extension of boosting.A cost-sensitive boosting learning framework is proposed based on cost-sensitive sampling.Through introducing cost-sensitive sampling in each round of naive boosting,the expectation of cost-sensitive loss is minimized.Under the above framework,two new cost-sensitive boosting algorithms are deduced.Meanwhile,issues of the instability existing in early cost-sensitive boosting algorithms are revealed and explained.Experimental results on UCI(University of California,Irvine)data set and CBCL(Center for Biological & Computational Learning)face data set demonstrate:in terms of the cost-sensitive classification problem,cost-sensitive sampling boosting algorithms outperform naive boosting and existing cost-sensitive boosting algorithms.