A Lightweight Ensemble Model for Real-Time Cyberbullying Detection in Low-Resource and Code-Mixed Contexts

Cyberbullying on social media platforms continues to escalate, particularly within multilingual and code-mixed digital ecosystems where low-resource languages such as Swahili remain significantly underrepresented. Existing cyberbullying detection systems are predominantly trained on high-resource languages like English and therefore suffer from reduced performance, poor cross-lingual generalization, and high computational demands when applied to linguistically diverse environments. Addressing these critical gaps, this study introduces a lightweight, energy-efficient Hybrid BLOOM-Based Ensemble Model (HBEM) tailored for real-time multilingual cyberbullying detection. The study also contributes two novel datasets, Swahili and Swanglish, collected from Twitter (currently X), enabling robust evaluation in both low-resource and code-mixed contexts. HBEM integrates multilingual large language models with parameter-efficient ensemble learning and optimized inference strategies to balance performance, scalability, and computational cost. The results demonstrate that HBEM outperforms state-ofthe-art models including BERT, RoBERTa, and mBERT across all evaluation metrics, with only a slightly higher energy consumption in certain configurations. The model achieves 97.86% accuracy, 97.83% precision, 97.21% recall, a 97.48% F1-score, an inference time of 13.1 ms, and an energy consumption of 0.001079 kWh, establishing new benchmarks for multilingual cyberbullying detection. These findings further highlight the strategic value of code-mixing, as Swanglish data improved crosslingual transfer and enhanced detection performance in lowresource Swahili settings. The study offers significant implications for researchers, policymakers, and industry practitioners seeking scalable, socially inclusive, and energy-efficient safety solutions for digital ecosystems. Overall, this study delivers a high-performance, linguistically adaptive, and computationally sustainable approach for strengthening cyberbullying mitigation across global social media platforms.