DocumentCode :
479096
Title :
Cost-Sensitive Parameters Estimation Method of Bayesian Networks
Author :
Gao, Yanfang ; Wang, Zhenzhen ; Chen, Yingwu
Author_Institution :
Coll. of Inf. Syst. & Manage., Nat. Univ. of Defense Technol., Changsha
fYear :
2008
fDate :
12-14 Oct. 2008
Firstpage :
1
Lastpage :
4
Abstract :
Misclassiflcation cost is usually unequal for different class. Parameters estimation methods of Bayesian networks such as generative method based on log joint likelihood loss and discriminative method based on log conditional likelihood loss suppose that misclassification cost of every class is equal. Accordingly, this paper develops a cost-sensitive parameters estimation method so that Bayesian networks can be used for cost sensitive classification. A cost sensitive loss function and an extended cost sensitive loss function are defined based on classification cost matrix and extended classification cost matrix respectively. The two cost-sensitive loss functions are both extension of log conditional likelihood loss in form. In experiment cost-sensitive Bayesian networks are compared with the corresponding generative Bayesian networks and discriminative Bayesian networks. Experiment results show that cost sensitive Bayesian networks with extended cost sensitive loss has preferable classification performance when used for cost sensitive classification.
Keywords :
belief networks; learning (artificial intelligence); maximum likelihood estimation; parameter estimation; Bayesian networks; discriminative method; extended classification cost matrix; extended cost sensitive loss function; generative method; log conditional likelihood loss; log joint likelihood loss; parameters estimation method; Bayesian methods; Closed-form solution; Cost function; Educational institutions; Information management; Learning systems; Management information systems; Parameter estimation; Performance loss; Technology management;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Wireless Communications, Networking and Mobile Computing, 2008. WiCOM '08. 4th International Conference on
Conference_Location :
Dalian
Print_ISBN :
978-1-4244-2107-7
Electronic_ISBN :
978-1-4244-2108-4
Type :
conf
DOI :
10.1109/WiCom.2008.2712
Filename :
4680901
Link To Document :
بازگشت