Detecting Cyberbullying using Machine Learning Approaches
Adnan Haider, Abu Bakar Siddique, Raja Hashim Ali, Muhammad Imad, Ali Zeeshan Ijaz, Usama Arshad, Nisar Ali, Memoona Saleem, Nazia Shahzadi
Abstract
Social media platforms serve as the main conduit for information and communication in today’s society. Social media platforms have ingrained themselves into our daily lives, and as access is expanded to more distant regions, their user bases are rapidly growing. Roman Urdu is the primary form of communication on social media among the 71.70 million users in Pakistan. As a result of these advancements and the rise in users, cyberbullying—also known as digital bullying—has increased. This study emphasizes on social media users who interact using Roman Urdu, an Urdu dialect written using the English alphabet. In this study, we investigated the issue of online bullying behavior on the data (Roman Urdu) that was collected from Kaggle. This is one of the rare research projects that, as far as we are aware, addresses Roman Urdu cyberbullying behavior. The purpose of the research we propose is to find a suitable model for Roman Urdu cyberbullying behavior classification. The data was in raw form in a way that the contents and data annotations were in the different lists inside different dictionaries of the root dictionary. After labeling the data, the dataset has then undergone pre-processing to remove duplications, stop words, punctuation, and other sources of noise. Then, based on supervised learning, a set of different learning algorithms, including Support Vector Machine, Decision Tree, Random Forest, Naïve Bayes, and Logistic Regression were applied to both types of extracted features. The features are extracted using two separate approaches, Count-Vectorizer and TF-IDF (term frequency-inverse document frequency) Vectorizer. With performance rates of 89.9% when applied to TF-IDF features and 88.3 percent when applied to CV features, Random Forest (RF) outperformed the other developed algorithms by both combinations. The suggested approach might assist online social apps and chat rooms create stronger bully word filters and make the internet a safer place for users.