Discriminatively learning selective averaged one-dependence estimators based on cross-entropy method

Abstract

Averaged One-Dependence Estimators [1], simply AODE, is a recently proposed algorithm which weakens the attribute independence assumption of naïve Bayes by averaging all the probability estimates of a collection of one-dependence estimators and demonstrates significantly high classification accuracy. In this paper, we study the selective AODE problem and proposed a Cross-Entropy based method to search the optimal subset over the whole one-dependence estimators. We experimentally test our algorithm in term of classification accuracy, using the 36 UCI data sets recommended by Weka, and compare it to C4.5[5], naïve Bayes, CL-TAN[6], HNB[7], AODE and LAODE[3]. The experiment results show that our method significantly outperforms all the other algorithms used to compare, and remarkably reduces the number of one-dependence estimators used compared to AODE. © Springer-Verlag Berlin Heidelberg 2007.