Feature selection techniques for gender prediction from blogs

摘要

The goal of this paper is to identify gender of blog authors. Features such as POS tags, unigram (words+punctuations), bigrams and word classes are considered. To synthesis/rank features we are using Mutual information, Chi-square and Information gain methods. The dataset is the collection of 3227 blogs originally derived from blogs set, and among them 1679 were written by male and 1548 were written by female. The results were obtained using 10-cross fold validation. Unigram of words gave better accuracy of 78.81% in comparison with the other features. We found that chi-square is the best in ranking features. The classification is done using Multinomial Naïve Bayes Classifier, and different kernel functions of SVM such as PolyKernel, Puk, Normalized PolyKernel and RBFkernel.