Fine tuning attribute weighted naive Bayes

Abstract

Naive Bayes (NB) is one of the top 10 data mining algorithms due to its simplicity, efficiency and efficacy. However, both the unrealistic attribute conditional independence assumption and the unreliable conditional probability estimation limit its performance. Of numerous improved approaches, attribute weighting only focuses on alleviating the unrealistic attribute conditional independence assumption, while fine tuning devotes all the efforts to finding a more reliable conditional probability estimation. In this study, we argue that both of them are equally important to enhance the performance of NB and propose a novel model called fine tuned attribute weighted NB (FTAWNB) by combining fine tuning with attribute weighting into a uniform framework. In FTAWNB, we first exploit correlation-based attribute weighting to initialize the conditional probabilities, then for each misclassified training instance, the conditional probabilities are fine tuned iteratively to make them more reliable, and the fine tuning process will stop once the training classification accuracy no longer improves. Extensive experimental results show that FTAWNB significantly outperforms all the other existing state-of-the-art competitors.