Extensions of Naive Bayes and Their Applications to Bioinformatics

摘要

In this paper we will study the naive Bayes, one of the popular machine learning algorithms, and improve its accuracy without seriously affecting its computational efficiency. Naive Bayes assumes positional independence, which makes the computation of the joint probability value easier at the expense of the accuracy or the underlying reality. In addition, the prior probabilities of positive and negative instances are computed from the training instances, which often do not accurately reflect the real prior probabilities. In this paper we address these two issues. We have developed algorithms that automatically perturb the computed prior probabilities and search around the neighborhood to maximize a given objective function. To improve the prediction accuracy we introduce limited dependency on the underlying pattern. We have demonstrated the importance of these extensions by applying them to solve the problem in discriminating a TATA box from putative TATA boxes found in promoter regions of plant genome. The best prediction accuracy of a naive Bayes with 10 fold cross validation was 69% while the second extension gave the prediction accuracy of 79% which is better than the best solution from an artificial neural network prediction.