A Hybrid Ensemble Framework for Cyberbullying Detection using Multi-Model Consensus and Confidence Weighting

Abstract

Cyberbullying detection in social media has becomea critical research challengedueto therapidgrowth ofonlinecommunicationplatforms.Traditionalapproachesrelyeitheronrule-basedsystemsordeeplearningmodels, each with inherent limitations such as poor generalization or lack of interpretability. This paper proposes a hybrid ensemble framework that integrates multiple decision engines, including rule-based logic and transformer-based models,combinedwithaconfidence-awareaggregationmechanism.AnovelMulti-EngineCyberbullyingFramework (MECF), along with a Multi-Class Weighted Grading (MCWG) strategy, is introduced to improve detection robustness. The system evaluates predictions from multiple models and aggregates them using confidence-based weighted voting to produce the final classification. Experimental results on a balanced dataset demonstrate that the proposed approach achieves anF1-score of 0.91 and an AUC score of 0.962, outperforming individual models. The results highlight the effectiveness of combining heterogeneous models with confidence-based consensus for robust cyberbullying detection.

Introduction

The paper presents an AI-based system for cyberbullying detection, prevention, and response to address the growing problem of harmful online behavior on social media platforms. Traditional methods such as keyword filtering and rule-based systems are ineffective at detecting contextual, sarcastic, or implicit abusive content, while even single deep learning models often struggle with noisy and diverse social media text.

To overcome these issues, the study proposes a hybrid ensemble framework (MECF) that combines rule-based systems with transformer-based models. It introduces a Multi-Class Weighted Grading (MCWG) mechanism that assigns confidence-based weights to each model’s prediction, allowing more reliable and balanced final decisions.

The system pipeline includes data preprocessing (cleaning text, tokenization, handling imbalance), feature extraction (BERT embeddings, sentiment and linguistic features), and classification using a fine-tuned DeBERTa model along with other engines. A dataset of over 5000 social media samples (Twitter, Reddit, YouTube, Kaggle) is used for training and evaluation.

Results show strong performance, with the proposed hybrid model achieving an AUC of 0.962 and outperforming traditional ML models (SVM, LSTM) and other transformers (BERT, RoBERTa). Overall, the system improves detection accuracy, robustness, and real-time intervention capability through prevention alerts and automated content moderation actions.

Conclusion

This paper presenteda hybrid ensembleframeworkfor cyberbullying detection that integrates rule-based and transformer-based models. The proposed MCWG- based decision mechanism enables effective aggregation of multiple predictions using confidence scores. Experimental results demonstrate that the hybrid approachimproves detectionperformancecomparedto individual models. The framework provides a balance betweenaccuracyandrobustness,makingitsuitablefor real-world applications. Future work includes incorporating sarcasm detection, improvingpreprocessingtechniques,andextendingthe model to multilingual datasets.

References

[1] T.Davidsonetal.,“AutomatedHateSpeech Detection,” ICWSM, 2017. [2] Z.WaseemandD.Hovy,“HatefulSymbolsor Hateful People,” NAACL, 2016. [3] J.Devlinetal.,“BERT:Pre-trainingofDeep Bidirectional Transformers,” NAACL, 2019. [4] Y.Liuetal., “RoBERTa:ARobustlyOptimized BERT Approach,” 2019. [5] P.Heetal.,“DeBERTa:Decoding-enhanced BERT,” ICLR, 2021. [6] T.Wolfetal.,“Transformers:State-of-the-Art NLP,” EMNLP, 2020. [7] F.Pedregosaetal.,“Scikit-learn:MachineLearning in Python,” JMLR, 2011. [8] I.Goodfellowetal.,DeepLearning,MITPress, 2016. [9] S.HochreiterandJ.Schmidhuber,“LongShort- Term Memory,” Neural Computation, 1997. [10] A.Vaswani etal., “Attention isAll YouNeed,” NeurIPS, 2017. [11] N.Vidgenetal.,“ChallengesinHateSpeech Detection,” 2019. [12] H.Zhangetal.,“DetectingOffensiveLanguagein Social Media,” ACL, 2018. [13] K.Chawlaetal.,“SMOTE:SyntheticMinority Over-sampling,” 2002. [14] J. Brownlee, “Imbalanced Classification,” Machine Learning Mastery, 2020. [15] S. Sun et al., “A Survey of Ensemble Methods,” Information Fusion, 2020. [16] KaggleDataset:CyberbullyingClassification Dataset, Available: https://www.kaggle.com

Copyright

Copyright © 2026 Atharva Kadam, Utkarsh Chavan, Ganesh Nagare, Kaushik Moger. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.