贝叶斯网络生成学习和判别学习对比研究

Comparative research on generative learning and discriminative learning of Bayesian networks

优化目标决定了贝叶斯网络分类器的分类性能。文章围绕生成函数和判别函数等两类典型的优化目标,对比分析了贝叶斯网络在不同学习目标下的学习方法,应用UCI数据集,通过实验对比了训练样本数量的变化对贝叶斯网络分类器性能的影响,分析了贝叶斯网络分类器的目标函数与分类性能的关系。数据实验结果表明:冗余数据对判别贝叶斯网络过拟合的影响大于生成贝叶斯网络,"最优"贝叶斯网络分类器并不一定具有最大的联合似然值或者条件似然值;为了提高学习效率和分类性能,可在训练判别贝叶斯网络的过程中采用主动样本选择策略,并且以生成函数和判别函数的权衡值作为贝叶斯网络分类器的优化目标。

Optimization function determines classification performance of Bayesian networks classifier. Different training methods based on generative function and discriminative function are compared, effect of increasing number of training samples on classification performance of generative Bayesian networks and discriminative Bayesian networks are compared and correlation between optimization function and classification performance of Bayesian networks are analyzed according to the experiment on UCI datasets. The experimental results show that redundant data have greater impact on over fitting in discriminative Bayesian networks than that of generative networks and the best Bayesian networks classifier do noalways have the biggest log likelihood or log conditional likelihood. So, adopting active sample selection strategy and taking tradeoff between generative and discriminative function as optimization objective may improve learning efficiency and classification performance of discriminative Bayesian networks.