基于L1正则化的贝叶斯网络分类器

目前基于节点排序的贝叶斯网络分类器忽略了节点序列中已选变量和类标签之间的信息,导致分类器的准确率很难进一步提高。针对这个问题,提出了一种简单高效的贝叶斯网络分类器的学习算法:L1正则化的贝叶斯网络分类器(L1-BNC)。通过调整Lasso方法中的约束值,充分利用回归残差的信息,结合点序列中已选变量和类标签的信息,形成一条优秀的有序变量拓扑序列(L1正则化路径);基于该序列,利用K2算法生成优良的贝叶斯网络分类器。实验表明,L1-BNC在分类精度上优于已有的贝叶斯网络分类器。L1-BNC也与SVM,KNN和J48分类算法进行了比较,在大部分数据集上,L1-BNC优于这些算法。

Bayesian Network Classifier Based on L1 Regularization

Variable order-based Bayesian network classifiers ignore the information of the selected variables in their sequence and their class label, which significantly hurts the classification accuracy. To address this problem, we proposed a simple and efficient Ll regularized I3ayesian network classifier (Ll-I3NC). Through adjusting the constraint value of Lasso and fully taking advantage of the regression residuals of the information, L1-BNC takes the information of the sequence of selected variables and the class label into account, and then generates an excellent variable ordering sequence(L1 regularization path) for constructing a good Bayesian network classifier by the K2 algorithm. Experimental results show that L1-BNC outperforms existing state-of-the-art Bayesian network classifiers. In addition, in comparison with SVM,Knn and J48 classification algorithms,L1-BNC is also superior to those algorithms on most datasets.