Comparative analysis on feature selection based Bayesian text classification

Feature selection is an important preprocessing step for data in the classification and regression learning. Many feature selection algorithms have been proposed using the different information criteria based on mutual information. However, there is no such comparative study conducted to analyse the effectiveness of these methods under a specific application framework. In this paper, we select 6 different feature selection algorithms, i.e, RelFss, MIFS-U, FCBF, CMIM, mRMR, and mMIFS-U, to compare their reduction capabilities and classification performances in the application of naive Bayesian based text classification. We collect a lot of documents belonging to ten different domains from the Chinese News Web site (www.people.com.cn) as the experimental data, where each of documents includes 1,000 Chinese characters at least. From the experimental results, we can conclude that naive Bayesian with the features selected by mRMR can obtain the highest classification accuracy. The summarized conclusions give some guidelines for feature selection in text classification application.