摘要
建立了具有数据缺失训练集下学习贝叶斯网的一种混合启发方法:SGS-EM-PACOB算法。它基于打分-搜索方法,利用GS和EM数据补全策略分别得到学习所需要的统计因子,并将两者联合起来作为PACOB算法的启发因子。实验证明,SGS-EM-PACOB算法充分保留GS和EM两者的优点,促使算法能够平稳地收敛到理想结果。相对于只具有单一数据补全策略的算法,该算法不仅在度量数据拟合程度的Logloss值上保持稳定,而且在学习到的贝叶斯网络结构上也有改进。
Presented an efficient hybrid heuristic SGS-EM-PACOB algorithm for learning Bayesian network with missing values. It is based on scoring and searching method by using GS and EM data completion policies to attain statistic information, which is essential in learning Bayesian network. SGS-EM-PACOB algorithm combines these two policies for PACOB, an excellent parallel ant colony heuristic for learning bayesian network with complete dataset. The experi- ments showed SGS-EM-PACOB algorithm fully out-performed both GS and EM, and made the algorithm converge to ideal results smoothly. Comparing with those algorithms having only one data completion policy, SGS-EM-PACOB algorithm not only achieves a stable Logloss value,which measures how well the dataset matches the learned network, but also makes improvements on the learned bayesian network structure.
出处
《计算机科学》
CSCD
北大核心
2008年第12期163-166,共4页
Computer Science
基金
国家教育部博士点基金(20060285008)
江苏省自然科学基金(BK2003030)资助
关键词
学习贝叶斯网
数据补全策略
混合启发
Learning bayesian network,Data completion policy, Hybrid heuristic