THE IMPROVEMENT OF NAIVE BAYESIAN CLASSIFIER BASED ON THE STRATEGY OF FEATURE SELECTION AND SAMPLE CLEANING

摘要

Naive Bayesian Classifier (NBC) is a simple and effective classification model. Though it shows a lot of edges over many other Classifiers, it does not always yield satisfactory result. In this paper, we give a summary of the previous improvement methods for the NBC model. Then three improvement strategies are proposed: the feature selection strategy, the sample cleaning strategy and the mixed strategy. By choosing the optimized feature subset according to the feature important factor (FIF) of every feature, the first method simplifies the dimensionality of dataset; the second method deletes noisy samples within the training dataset according to the sample polluting factor; the third method integrates the two methods: first the feature selection, and then the sample cleaning. Through the experimental comparison and analysis on the UCI repository, these strategies are proved effective. Averagely speaking, with 36.76% of the features in the original feature set, we can raise the prediction accuracy by 2.30% using the first method. While with 92.57% samples in the training dataset, we can raise the prediction accuracy by 1.59% using the second method. As to the third method, the prediction accuracy can be increased 2.55%. Among these strategies, the mixed one shows the advantages over the other two, which reduces the complexity of the model while increasing the prediction accuracy of the NBC model.