Abstract
The widespread usage of social media has led to the increasing popularity of online advertisements, which have been accompanied by a disturbing spread of clickbait headlines. Clickbait dissatisfies users because the article content does not match their expectation. Detecting clickbait posts in online social networks is an important task to fight this issue. Clickbait posts use phrases that are mainly posted to attract a user's attention in order to click onto a specific fake link/website. That means clickbait headlines utilize misleading titles, which could carry hidden important information from the target website. It is very difficult to recognize these clickbait headlines manually. Therefore, there is a need for an intelligent method to detect clickbait and fake advertisements on social networks. Several machine learning methods have been applied for this detection purpose. However, the obtained performance (accuracy) only reached 87% and still needs to be improved. In addition, most of the existing studies were conducted on English headlines and contents. Few studies focused specifically on detecting clickbait headlines in Arabic. Therefore, this study constructed the first Arabic clickbait headline news dataset and presents an improved multiple feature-based approach for detecting clickbait news on social networks in Arabic language. The proposed approach includes three main phases: data collection, data preparation, and machine learning model training and testing phases. The collected dataset included 54,893 Arabic news items from Twitter (after pre-processing). Among these news items, 23,981 were clickbait news (43.69%) and 30,912 were legitimate news (56.31%). This dataset was pre-processed and then the most important features were selected using the ANOVA F-test. Several machine learning (ML) methods were then applied with hyper-parameter tuning methods to ensure finding the optimal settings. Finally, the ML models were evaluated, and the overall performance is reported in this paper. The experimental results show that the Support Vector Machine (SVM) with the top 10% of ANOVA F-test features (user-based features (UFs) and content-based features (CFs)) obtained the best performance and achieved 92.16% of detection accuracy.