FWIB: Fuzzy Weighted Iterative Bayes Algorithm

Abstract

Within the theoretical framework of Naive Bayes(NB), Iterative Bayes(IB), as an important variant, was proposed to alleviate the attribute interdependence problem based on output adjustment technique and shows certain performance enhancement. However, like NB, IB also suffers from biased classification and poor robustness against noise on class imbalanced data. To address these problems, cost-sensitive learning is introduced, and by assigning fuzzy-membership values to training data, we propose a class of Fuzzy Weighted Iterative Bayes algorithms (FWIB-CE, FWIB-HE, FWIB-CL-HL) to integrate more prior information from the sample distribution. The experimental results on ten typical imbalanced datasets show our methods have significant outperformance over NB and IB under G-mean evaluation in classification and have higher robustness against noise interference.