摘要
针对最小化错误分类器不一定满足最小化误分类代价的问题,提出了一种代价敏感准则——即最小化误分类代价和最小化错误分类率的双重准则.研究了基于代价敏感准则的贝叶斯网络结构学习,要求搜索网络结构时在满足误分类代价最小的同时,还要满足错误分类率优于当前的最优模型.在UCI数据集上学习代价敏感贝叶斯网络,并与相应的生成贝叶斯网络和判别贝叶斯网络进行比较,结果表明了代价敏感贝叶斯网络的有效性.
Classifier based on minimum classification error may not minimize classification cost. A cost-sensitive model selection criterion is proposed. This criterion is in fact a dual rule which fulfills minimum classification cost and minimum classification error. Bayesian network structure learning based on the cost sensitive criterion is researched. In the procedure of search potential structure the criterion demands that the selected model must fulfill minimum classification cost and lower classification error than the current best model. Cost-sensitive Bayesian networks are learned from dataset of UCI and then are compared with the corresponding generative Bayesian networks and discriminative Bayesian networks, results show that cost-sensitive Bayesian network is effective.
出处
《小型微型计算机系统》
CSCD
北大核心
2009年第2期313-316,共4页
Journal of Chinese Computer Systems
基金
高等学校博士学科点专项科研基金项目(20059998019)资助
关键词
贝叶斯网络
误分类代价
代价敏感准则
判别准则
生成准则
Bayesian networks
classification error cost
cost-sensitive eriterion
discriminative criterion
generative criterion
