Hybrid 1D VGG16 and SVM Framework for Early Detection and Intensity Classification of Cyberbullying

Cyberbullying is a growing concern on social media, with severe psychological and emotional consequences.This study proposes a hybrid 1D VGG16 + SVM framework for cyberbullying detection and intensity classification, integrating deep learning-based feature extraction with SVM's classification capabilities.The model, trained on the Kaggle Cyberbullying Dataset (47,733 tweets), classifies cyberbullying into low, medium, and high-intensity levels to enable targeted interventions.The proposed model achieves a high accuracy of 97.86%, surpassing state-of-the-art approaches such as Obaid's LSTM + Fuzzy Logic (93.67%),Alqahtani et al.'s Ensemble Learning (90.71%), and Ogunleye et al.'s BERT-based method (96.0%).Its ability to detect early signs of cyberbullying and classify severity levels ensures effective intervention at various stages of escalation.By combining 1D VGG16 for feature extraction with SVM for precise classification, the model optimally balances accuracy and computational efficiency, reducing training time (367.93s)compared to 1D VGG16 alone (753.21s).SHAP-based interpretability ensures transparency by highlighting the impact of sentiment polarity, profanity count, and aggression score in cyberbullying detection.SMOTE resampling improves minority-class detection to tackle class imbalance, while intensity-based categorization enhances intervention strategies.Future research will explore multilingual datasets and real-time adaptive detection systems to improve scalability and generalization.This study establishes a scalable, efficient, and interpretable framework for cyberbullying detection, contributing to safer online environments.