Improved Hybrid Model for Robust Cyberbullying Detection
Promise A. Nlerum, Beauty Brisibe · INTERNATIONAL JOURNAL OF MATHEMATICS AND COMPUTER RESEARCH · 2024
Cyberbullying, defined as the intentional and repeated harassment through digital platforms such as social media, has become a growing concern due to its severe impact on mental health, including anxiety, depression, and social isolation. The increasing prevalence of such behaviors highlights the urgent need for effective detection systems to mitigate harm and foster safer online interactions. This study investigates a hybrid model for robust cyberbullying detection on social media, specifically Twitter (now X), by combining unsupervised learning techniques, sentiment analysis, and Long Short-Term Memory (LSTM) networks for accurate text classification. The model's performance was evaluated using the Russian Troll dataset, focusing on key metrics such as accuracy, precision, recall, and F1-score. Sentiment analysis, an essential task in natural language processing (NLP), classifies text into positive, negative, or neutral categories by extracting underlying emotions. The hybrid model integrates tokenization, Term Frequency-Inverse Document Frequency (TF-IDF) for feature extraction, and LSTM layers for sequential modeling. The results demonstrate the model's effectiveness, achieving an impressive 94.9% accuracy, with precision, recall, and F1-score all reaching 95%. These findings highlight its ability to minimize false positives and negatives while handling the noisy, ambiguous nature of social media content. The study underscores the model's robustness in sentiment classification, positioning it as a powerful tool for detecting cyberbullying and harmful online interactions.