Non-linguistic Features for Cyberbullying Detection on a Social Media Platform Using Machine Learning

Abstract

Cyberbullying on social media platforms has been a severe problem with serious negative consequences. Therefore, a number of researches on automatic detection of cyberbullying using machine learning techniques have been conducted in recent years. While cyberbullying detection has traditionally utilized linguistic features, the cyberbullying on social media does not have only linguistic features. In this paper, a holistic multi-dimensional feature set is developed which takes into account individual-based, social network-based, episode-based and linguistic content-based cyberbullying features. To test performance of the proposed multi-dimensional feature set, we designed and built cyberbullying detection models on the KNIME machine learning platform. Six different machine learning algorithms - Naive Bayes, Decision Tree, Random Forest, Tree Ensemble, Logistic Regression, and Support Vector Machines - were used in our cyberbullying detection models. Our experimental results demonstrate that applying the proposed multi-dimensional feature set (i.e. the set not limited to the linguistic features) results in an improved cyberbullying detection for all tested machine learning algorithms.