What is it about?

Cyberbullying is a growing concern on social media, yet AI detection systems often struggle to recognize harmful messages that depend on context or contain subtle expressions. Our research introduces **CyDeBoost**, an AI framework that improves cyberbullying detection by learning from its own mistakes. The framework uses explainable AI (SHAP) to identify the words and language patterns associated with incorrect predictions. It then uses large language models to generate additional training examples that help the detection system better distinguish harmful from harmless content. By combining these techniques with continual learning, CyDeBoost improves detection performance while preserving previously acquired knowledge. This research offers a practical approach to developing more accurate, adaptable, and explainable AI systems for safer online communication.

Featured Image

Why is it important?

Cyberbullying is increasingly difficult to detect because harmful messages can be subtle, sarcastic, or highly dependent on context. Even advanced AI systems may misinterpret these expressions, allowing harmful content to go unnoticed. Our study introduces CyDeBoost, a framework that helps AI systems learn from their mistakes. It identifies why certain messages were misclassified and uses these insights to generate targeted training examples. The system then improves its ability to recognize cyberbullying while retaining previously learned knowledge. Evaluations across three social media datasets demonstrate improvements in detection performance. These findings offer a promising direction for developing more accurate, explainable, and adaptable content moderation systems that can help online platforms identify harmful interactions more effectively.

Perspectives

As a researcher working in artificial intelligence and online safety, I have often wondered how we can help AI systems learn more effectively from their mistakes. Developing CyDeBoost gave me an opportunity to explore this question. One of the most interesting aspects of this research was discovering how explanations of incorrect predictions could guide a model toward better decisions. It also reinforced my belief that understanding why an AI system makes mistakes is an important part of improving its reliability. I hope this work encourages researchers to explore AI systems that can continuously learn from their errors while preserving what they have already learned. Ultimately, I would like to see these ideas contribute to safer online environments and more trustworthy AI technologies.

Shih-Wei Guo
Feng Chia University

Read the Original

This page is a summary of: CyDeBoost: A SHAP-Guided Boosting Strategy Framework for Cyberbullying Detection through Mistake-Driven Learning, ACM Transactions on the Web, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3831585.
You can read the full text:

Read

Contributors

The following have contributed to this page