Feature ranking for Persian Review Spam detection

Using online reviews is one of the main factors in customers’ decision making for buying a product or using a service. These reviews are valuable sources of information which can be used for detecting public opinion about products or services. Although online reviews are useful, trusting them blindly is dangerous for both costumers and sellers as they may be manipulated to earn profit; such reviews are called spam reviews. The current study addresses Persian reviews about cell-phone extracted from Digikala.com and investigates spam type 1 and type 2 which are fake reviews and reviews describing brands’ names only, respectively. Features used in this study, due to their efficiency, are review-based and metadata features. These features and their combinations in detecting Persian spam reviews, also their effect on the accuracy of classifier are assessed. Spam classification is performed using decision tree, support vector machines, and naïve Bayes classifiers and their accuracy are compared using different features’ combinations. The highest accuracy is obtained using the decision tree classifier which achieves 0.778 in terms of F-measure. In ranking features, again the decision tree outperforms the other two classifiers by achieving 0.824 F-measure by combining the positive feedback, overall score, and review polarity features.