摘要
为了测试评估贝叶斯分类器的性能,用不同数据集进行对比实验是必不可少的.现有的贝叶斯网络实验软件包都是针对特定目的设计的,不能满足不同研究的需要.介绍了用Matlab在BNT软件包基础上建构的贝叶斯分类器实验平台MBNC,阐述了MBNC的系统结构和主要功能,以及在MBNC上建立的朴素贝叶斯分类器NBC,基于互信息和条件互信息测度的树扩展的贝叶斯分类器TANC,基于K2算法和GS算法的贝叶斯网络分类器BNC.用来自UCI的标准数据集对MBNC进行测试,实验结果表明基于MBNC所建构的贝叶斯分类器的性能优于国外同类工作的结果,编程量大大小于使用同类的实验软件包,所建立的MBNC实验平台工作正确、有效、稳定.在MBNC上已经进行贝叶斯分类器的优化和改进实验,以及处理缺失数据等研究工作.
To test and evaluate the performance of Bayesian Classifier, it is absolutely necessary to carry through contrastive experiment using different data sets. Current packages for Bayesian Classifier experiment are designed for certain purposes, so that it can't satisfy the needs of different research. It introduces the building of experiment platform MBNC for Bayesian Classifiers using Matlab based on BNT, including the system structure and the main function of MBNC, the Classifiers built on MBNC: the Nave Bayesian Classifier NBC, the Tree Augmented Nave Bayesian Classifier TANC based on Mutual Information and Conditional Mutual Information, Bayesian Network Classifier BNC based on K2 and GS algorithm. MBNC is tested by standard data set from UCI and the results show that the performance of Bayesian classifiers built on MBNC are preceded similar works and the quantity of programming much less than that using current packages, which indicates that the platform works correctly, effectively and stably. Now the experiments for optimizing Bayesian Classifiers and the study of dealing with missing data are carried through on MBNC.
出处
《复旦学报(自然科学版)》
CAS
CSCD
北大核心
2004年第5期729-732,共4页
Journal of Fudan University:Natural Science
基金
清华大学智能技术与系统国家重点实验室开放课题项目资助(99002)