Enhance Social Network Bullying Detection using Transfer Learning with Random Forest Classifier

Cyberbullying involves the use of digital platforms to inflict emotional or psychological harm on individuals through activities like posting insulting messages, images, or videos with the intent to humiliate, harass, or threaten. This harmful behavior can quickly spread, causing victims to experience regret, humiliation, or defamation. In Thailand, cyberbullying is a prevalent issue, particularly among teenagers using mobile technology. To mitigate the risk for potential victims, there is a need for a tool capable of identifying cyberbullying behaviors. However, existing research on this topic is limited. This study proposes a solution by developing an automated system that detects cyberbullying traces on social media, utilizing natural language processing techniques to identify and prevent such incidents. In this research, we combining BERT (Bidirectional Encoder Representations from Transformers) and Decision Forests for cyberbullying detection, which involves fine-tuning a pretrained BERT model as input features for a Decision Forests classifier. The experimental outcomes suggest that integrating BERT (Bidirectional Encoder Representations from Transformers) with Random Forests for cyberbullying detection, with the fine-tuning of a pre-trained BERT model as input features for a Random Forests classifier, is successful for training and determining the type of cyberbully in social media. The proposed method achieves the highest performance across all datasets, with accuracy values of 0.925, 0.905, and 0.910 in Wisesight dataset, Thai Toxic Tweet dataset, and The 40 Thai Children Stories dataset in classifying the type of social media cyberbully.