Abstract
Due to the proliferation of internet-enabled smartphones, many people, particularly young people in Arabic society, have widely adopted social media platforms as a primary means of communication, interaction and friendship making. The technological advances in smartphones and communication have enabled young people to keep in touch and form huge social networks from all over the world. However, such networks expose young people to cyberbullying and offensive content that puts their safety and emotional well-being at serious risk. Although, many solutions have been proposed to automatically detect cyberbullying, most of the existing solutions have been designed for English speaking consumers. The morphologically rich languages-such as the Arabic language-lead to data sparsity problems. Thus, render solutions developed for another language are ineffective once applied to the Arabic language content. To this end, this study focuses on improving the efficacy of the existing cyberbullying detection models for Arabic content by designing and developing a Consensus-based Ensemble Cyberbullying Detection Model. A diverse set of heterogeneous classifiers from the traditional machine and deep learning technique have been trained using Arabic cyberbullying labeled dataset collected from five different platforms. The outputs of the selected classifiers are combined using consensus-based decisionmaking in which the F1-Score of each classifier was used to rank the classifiers. Then, the Sigmoid function, which can reproduce human-like decision making, is used to infer the final decision. The outcomes show the efficacy of the proposed model comparing to the other studied classifiers. The overall improvement gained by the proposed model reaches 1.3% comparing with the best trained classifier. Besides its effectiveness for Arabic language content, the proposed model can be generalized to improve cyberbullying detection in other languages.