Feature selection for author identification of Persian online short texts

The growing use of social media and online communication to express opinions, exchange ideas, and also the expanding use of of this platforms by Persian users has increased Persian texts on the Web. This remarkable growth, along with abusive use of the writer's anonymity, reveals the need for the author's automatic identification system in this language. In this research, the purpose of the study is to investigate the factors affecting the identification of authors of Persian reviews produced by cell-phone buyers and also to evaluate supervised and unsupervised methods. The factors considered in this research include lexical, syntactic, semantic, structural, grammatical, text-specific, and specific to social networks. After extracting these features, selecting the best features is tested by four algorithms including feature correlation, gain ratio, OneR, and principal components analysis. In the following, K-means, EM and density-based clustering will be used for clustering and Bayesian network, random forest, and Bagging will be used for categorization. The evaluation of the above algorithms on Persian comments of Samsung phone buyers indicates that the best performance among the clustering algorithms is 59/16% obtained by the EM algorithm on top-15 features selected by OneR, while the random forest algorithm using top-90 features selected by gain ratio with 79/57% achieves the best performance among the classification algorithms. Also, the comparison of features showed that syntactic features had the most effect on the identification of the author of short texts, and then, lexical, text-specific, specific to social networks, structural, grammatical and semantic features, respectively.