Most of the data mining and machine learning algorithms will work better with discrete data rather than continuous. But the real time data need not be always discrete and thus it is necessary to discretize the continuous features. There are several discretization methods available in the literature. This paper compares the two methods Median Based Discretization and ChiMerge discretization. The discretized values obtained using both methods are used to find the feature relevance using Information Gain. Using the feature relevance, the original features are ranked by both methods and the top ranked attributes are selected as the more relevant ones. The selected attributes are then fed into the Naive Bayesian Classifier to determine the predictive accuracy. The experimental results clearly show that the performance of the Naive Bayesian Classifier has improved significantly for the features selected using Information Gain with Median Based Discretization than Information Gain with ChiMerge discretization.

Feature selection is an indispensable pre-processing technique for selecting more relevant features and eradicating the redundant attributes. Finding the more relevant features for the target is an essential activity to improve the predictive accuracy of the learning algorithms because more irrelevant features in the original feature space will cause more classification errors and consume more time for learning. Many methods have been proposed for feature relevance analysis but no work has been done using Bayes Theorem and Self Information. Thus this paper has been initiated to introduce a novel integrated approach for feature weighting using the measures viz., Bayes Theorem and Self Information and picks the high weighted attributes as the more relevant features using Sequential Forward Selection. The main objective of introducing this approach is to enhance the predictive accuracy of the Naive Bayesian Classifier.