贝叶斯网络分类器结构与变量分布的差异性分析

摘要

To improve the classification accuracy by discriminative learning strategy, we analyze the performance of discriminative parameter learning strategy with different restrictive Bayesian networks.In this experiment, we build a new structure by deleting edges on a tree structure by partial derivatives of log conditional likelihood. The results show that the discriminative strategy runs well when the structure is simpler than the truth, and reduces performance when there are redundant edges. These results change the recognition that the redundant edges are irrelevant to classification performance.%为了提高鉴别式学习策略训练的贝叶斯网络分类器的分类精度,分析了贝叶斯网络结构与数据中变量分布之间的差异对贝叶斯网络分类器性能的影响,实验以网络结构的实际联合概率分布的树型近似描述为基准,删除在条件对数似然函数极大化过程中不起作用的边,生成具有同一联合概率分布的不同描述程度的网络结构.实验结果表明,只有当网络结构表现力不足时,鉴别式参数学习才能起积极作用;而当网络结构中有多余的边时,反而容易受其制约.从而验证了网络中多余的边对分类器性能没有影响的观点是片面的.