Filter-Wrapper Approach to Feature Selection Using PSO-GA for Arabic Document Classification with Naive Bayes Multinomial

Text categorization and feature selection are two of the many text data mining problems. In text categorization, the document that contains a collection of text will be changed to the dataset format, the dataset that consists of features and class, words become features and categories of documents become class on this dataset. The number of features that too many can cause a decrease in performance of classifier because many of the features that are redundant and not optimal so that feature selection is required to select the optimal features. This paper proposed a feature selection strategy based on Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) methods for Arabic Document Classification with Naive Bayes Multinomial (NBM). Particle Swarm Optimization (PSO) is adopted in the first phase with the aim to eliminate the insignificant features and prepared the reduce features to the next phase. In the second phase, the reduced features are optimized using the new evolutionary computation method, Genetic Algorithm (GA). These methods have greatly reduced the features and achieved higher classification compared with full features without features selection. From the experiment that has been done the obtained results of accuracy are NBM 85.31%, NBM-PSO 83.91% and NBM-PSO-GA 90.20%.