基于条件对数似然函数导数的贝叶斯网络分类器优化算法

通常基于鉴别式学习策略训练的贝叶斯网络分类器有较高的精度,但在具有冗余边的网络结构之上鉴别式参数学习算法的性能受到一定的限制.为了在实际应用中进一步提高贝叶斯网络分类器的分类精度,该文定量描述了网络结构与真实数据变量分布之间的关系,提出了一种不存在冗余边的森林型贝叶斯网络分类器及其相应的FAN学习算法(Forest-Augmented Naive Bayes Algorithm),FAN算法能够利用对数条件似然函数的偏导数来优化网络结构学习.实验结果表明常用的限制性贝叶斯网络分类器通常存在一些冗余边,其往往会降低鉴别式参数学习算法的性能；森林型贝叶斯网络分类器减少了结构中的冗余边,更加适合于采用鉴别式学习策略训练参数；应用条件对数似然函数偏导数的FAN算法在大多数实验数据集合上提高了分类精度.

An Optimization Algorithm of Bayesian Network Classifiers by Derivatives of Conditional Log Likelihood

In general,Bayesian network classifiers trained by discriminative strategy have higher classification accuracy than others.However,the performance of discriminative parameter learning algorithms is limited in dealing with redundant edges.In order to improve the classification accuracy in a real situation,in this paper we describe the quantitative relations between Bayesian network structures and joint probability distributions,propose a Forest-Augmented Nave Bayes classifier and its learning algorithm.An FAN classifier is a kind of Bayesian network whose structure has few redundant edges,and FAN algorithm is optimized by properties of partial derivatives of conditional log likelihood.Experimental results have shown that redundant edges in the structure of a Bayesian network classifier could degrade classification performance in common situation,and most of restricted Bayeisian network classifiers have redundant edges.Therefore,FAN classifier without redundant edges is suitable for discriminative parameter learning strategy.On most of datasets,classification accuracies of classifiers trained by FAN algorithm are enhanced.