Cyberbullying Messages Detection Using Machine Learning and Deep Learning

Abstract

Cyberbullying has emerged as a significant concern in contemporary times, particularly due to its severe consequences, especially for children. In this paper, we propose an innovative machine learning-based approach aimed at accurately detecting cyberbullying messages and mitigating their harmful effects. The primary objectives of our research were twofold: developing a model capable of precisely identifying cyberbullying messages while distinguishing them from regular messages. To achieve this, we utilized a dataset of social media messages, labeled as normal, offensive, or hate messages. We adapted this dataset for binary classification, differentiating between cyberbullying and non-bullying messages. Our approach involved two distinct methods: firstly, utilizing Term Frequency-Inverse Document Frequency (TF-IDF) for traditional machine learning algorithms, and secondly, embedding texts for deep learning algorithms. We employed a total of 15 classifiers and performes a comprehensive comparison. The most successful algorithms from the first method were combined into a voting classifier, which demonstrated the highest accuracy of 96.5% during testing. Additionally, we assessed the impact of Recursive Feature Elimination with Cross-Validation (RFECV) on the model's performance and compared it with our baseline approach. Although the results exhibited slight fluctuations, the voting classifier consistently outperformed others with 96.6% accuracy. Our findings underline the effectiveness of the voting classifier based on machine learning algorithms, which delivered the most promising results. This approach holds the potential to be implemented in social media platforms or chat applications, serving as a valuable tool in the ongoing efforts to combat cyberbullying.