Feature weighting for naïve Bayes using multi objective artificial bee colony algorithm

Abstract

Naive Bayes (NB) is a widely used classifier in the field of machine learning. However, its conditional independence assumption does not hold true in real-world applications. In literature, various feature weighting approaches have attempted to alleviate this assumption. Almost all of these approaches consider the relationship between feature-class (relevancy) and feature-feature (redundancy) independently, to determine the weights of features. We argue that these two relationships are mutually dependent and both cannot be improved simultaneously, i.e., form a trade-off. This paper proposes a new paradigm to determine the feature weight by formulating it as a multi-objective optimisation problem to balance the trade-off between relevancy and redundancy. Multi-objective artificial bee colony-based feature weighting technique for naive Bayes (MOABC-FWNB) is proposed. An extensive experimental study was conducted on 20 benchmark UCI datasets. Experimental results show that MOABC-FWNB outperforms NB and other existing state-of-the-art feature weighting techniques.