Performance Evaluation of Quantum Machine Learning Models for Cyberbullying Detection

Abstract

Cyberbullying on social media has become a serious issue that affects users' mental well-being, safety, and online engagement. Detecting cyberbullying automatically is challenging due to informal writing styles, short messages, and the presence of subtle abusive patterns that often overlap with normal communication. This paper presents a comparative evaluation of quantum-based learning models for cyberbullying detection using a unified experimental pipeline. The proposed system applies text preprocessing followed by TF-IDF feature extraction and PCA-based dimensionality reduction, and then performs classification using Quantum Support Vector Machine (QSVM), Quantum Neural Network (QNN), Quantum Natural Language Processing (QNLP), and Quantum Singular Value Decomposition (QSVD). In addition, a hybrid QNN+QSVM ensemble is evaluated to examine whether combining complementary decision mechanisms improves prediction consistency. Experimental results show that QSVM achieved the best overall accuracy of 88.15%, with 92.52% precision, 92.23% recall, and an F1-score of 92.38%. QSVD delivered comparable performance with 88.08% accuracy, 93.53% precision, 90.98% recall, and an F1-score of 92.24%, indicating strong class separation with reduced feature complexity. QNN and QNLP produced competitive results, achieving 84.12% and 83.85% accuracy, respectively, with F1-scores close to 90%, demonstrating stable detection capability under quantum embedding. The hybrid QNN+QSVM model achieved 87.37% accuracy and an F1-score of 91.87%, showing balanced precision (92.06%) and recall (91.68%) and confirming that ensemble integration can improve reliability over individual quantum neural approaches. These findings suggest that kernel-driven quantum classifiers and decomposition-based quantum representations can provide strong performance for cyberbullying detection while maintaining consistent predictive behavior across evaluation measures.