摘要
当训练数据充分时,极大似然估计方法是贝叶斯网络参数学习典型且有效的方法。但当训练数据量少且领域知识缺乏时,极大似然估计往往无法给出一致无偏的参数估计。为此,提出一种新的贝叶斯网络参数学习方法TL-WMLE。将极大似然估计方法与迁移学习理论、样本不均衡方法相结合,解决数据量过少、领域知识缺乏时的贝叶斯网络参数学习问题。使用SMOTE-N方法构建辅助分类器,并依据协变量偏移理论,利用辅助分类器的分类结果来计算源域数据权值。采用赋权的源域数据和目标域数据构造目标域的似然函数,应用该似然函数对目标域的参数进行极大似然估计。实验结果表明,在小样本情况下,该方法的分类精度优于极大似然估计方法。
Maximum likelihood estimation is a classical and effective method for Bayesian network parameter learning on large samples, but it is not consistent when learning on small sample with little expertise. To address the issue, a novel method called TL-WMLE is proposed for Bayesian network parameter learning, which combines maximum likelihood, transfer learning and imbalance sample methods. The novel method uses an auxiliary classifier constructed by the SMOTE-N method and covariate migration theory, and computes the weights of source samples according to the predicted probability of the source domain by the auxiliary classifier. Then the proposed method mixes the reweighted source train sample and the target train sample to build a likelihood function on the target domain, and uses the new likelihood function to learn the parameters of the target domain via maximum likelihood estimation. Experimental results demonstrate that the classification accuracy of the proposed method outperforms that of the likelihood method on small samples.
出处
《计算机工程》
CAS
CSCD
北大核心
2016年第8期153-159,165,共8页
Computer Engineering
基金
国家自然科学基金资助项目"机器学习核方法模型选择与组合的核矩阵近似分析方法"(61170019)
关键词
贝叶斯网络
参数学习
小样本
迁移学习
目标域
Bayesian Network（BN）
parameter learning
small sample
transfer learning
target domain
