Preventing Digital Harm through AI: A Hybrid Model for Detecting Cyber-Bullying and Emotional Risk on Social Media

Abstract

The rise of social media has introduced a new form of psychological toxicity cyber-bullying, which significantly threatens mental well-being, particularly among adolescents and young adults. The emotional consequences of online harassment, such as anxiety, depression, and suicidal ideation, have increasingly been recognized as public health risks. The Hybrid Supervised Cyber-bully Detection System (HS-CBDS) integrates multiple machine learning techniques to enhance classification accuracy and reliability. The system incorporates comprehensive text preprocessing including normalization, tokenization, stopword removal, and lemmatization followed by feature engineering using TF-IDF, Bag of Words (BoW), sentiment polarity, and text length. A rule-based filter checks for offensive terms using a curated lexicon. The model employs supervised learning classifiers Linear SVC, Logistic Regression, and Random Forest optimized via hyperparameter tuning and GridSearchCV. In the HS-CBDS system, ensemble learning is implemented by combining predictions from five different classifiers LinearSVC, Logistic Regression, Random Forest, Decision Tree, and MLPClassifier. Each model's prediction is assigned an equal weight (0.2), and the final decision is made by averaging these outputs and rounding the result. Experimental results demonstrate that HS-CBDS outperforms individual classifiers, achieving an accuracy of 99.33%, significantly outperforming individual models, including ANN (98.17%) and Random Forest (97.99%). Evaluation metrics such as precision (99.34%), recall (99.33%), F1-score (99.33%), and ROC-AUC (0.9932) further validate the robustness and reliability of the system. HS-CBDS demonstrates a scalable, interpretable, and high-performing solution suitable for real-time cyber-bullying detection across social media and educational platforms, contributing effectively to a safer digital environment.