A Novel One-dependence Estimator Based on Multi-parents

Numerous approaches have been proposed to improve the classification accuracy of Naive Bayes by weakening the attribute independence assumption. To maintain the simple structure and low computational cost, many researches focus on the one-dependence estimator. In this paper, we present a novel algorithm called one-dependence estimator based on multi-parents (MPODE). In MPODE, each attribute has multi-parents; we use the mean dependence probability between the attribute and its parents as one dependence estimator. Experimentally testing on the whole 36 UCI datasets recommended by Weka, we compare our algorithm with NB, C4.5 by Quinlan (1993), SBC (selective Bayesian classifiers) by Langley and Sage (1994), TAN by Friedman et al., (1997) and AODE (aggregating one-dependence estimators) by Webb et al., (2005). The result shows that our algorithm outperforms NB, C4.5, SBC, and TAN significantly, and is almost same to AODE in term of classification accuracy