面向Stacking集成的改进分类算法及其应用

为了提高Stacking集成算法的分类性能,充分利用Stacking学习机制产生的先验信息和贝叶斯网络丰富的概率表达能力,提出一种基于属性值加权朴素贝叶斯算法的Stacking集成分类算法AVWNB-Stacking(Stacking based Attribute Value Weight Naive Bayes)。通过考虑属性值这个深层次的因素,以互信息(Mutual Information,MI)作为权值度量的基础,对属性权值向量横向扩展为每个属性值分配一个权值,避免不同的属性值共享相同的权值,从而解决朴素贝叶斯算法作为Stacking元分类器由于属性独立性假设带来的分类精度损失。实验结果表明,相比于传统算法及其他元分类器的Stacking分类算法,AVWNB-Stacking算法有效提高了模型的分类性能,在两个测试集上AUC值分别达到了0.8007和0.8607。

IMPROVED CLASSIFICATION ALGORITHM FOR STACKING INTEGRATION AND ITS APPLICATION

In order to improve the classification performance of the Stacking integration algorithm,making full use of the a priori information generated by learning mechanism of Stacking and the rich probability expression ability of the Bayesian network,a Stacking integrated classification algorithm based on attribute value weighted Naive Bayes algorithm,AVWNB-Stacking(Stacking based Attribute Value Weight Naive Bayes),is proposed.By considering the deep factor of attribute values and using mutual information(MI)as a basis of weight measure,we expanded horizontally the attribute weight vector and assigned a weight to each attribute value,avoiding different attribute values sharing the same weight value,thereby solving loss of classification accuracy brought by the Naive Bayes algorithm as a Stacking meta classifier due to attribute independence assumptions.The experimental results show that compared with the traditional algorithms and other meta-classifiers Stacking classification algorithm,the AVWNB-Stacking algorithm effectively improves the classification performance of the model,and the AUC value reaches 0.8007 and 0.8607 on the two test sets respectively.