Detecting Spam Reviews in Arabic by Deep Learning

Abstract

Online reviews are frequently used by consumers to make decisions about online purchases, hotel bookings, car rentals, and other choices because online shopping has grown in popularity over the past few years. Reviews are now crucial to both the customer and the business. As writing fake reviews comes with financial gain, opinion spam activities have increased. Some unethical companies may hire workers to write reviews to influence consumers’ purchasing decisions; therefore, detecting spam reviews is a very important task. We compiled a large dataset of Arabic reviews consisting of spam and non-spam that are categorized by crowd-sourcing approach. Then, we applied deep learning algorithms to detect spam reviews. To the best of our knowledge, there are no prior studies utilized deep learning to classify reviews that are written in Arabic. Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) models were used, and an accuracy of 97% was achieved by both algorithms. To further improve the results, unbalanced issues were solved by oversampling and undersampling techniques. The results of them are improvements in the precision, recall, and F1-score for spam reviews. For example, in CNN F1-score for spam class increased from 79% to 90% with undersampling and became 82% with oversampling.