<![CDATA[Cyberbullying has become a critical issue in social media use because it can negatively impact users’ mental health and social interactions. The high volume of aggressive comments and hate speech on digital platforms highlights the need for an automatic detection system that can accurately and reliably identify cyberbullying content. This research compares the performance of Indonesian language transformer models, IndoBERT and IndoBERTweet, in detecting text-based cyberbullying. Before modeling, the dataset undergoes Exploratory Data Analysis to understand its characteristics, class distribution, comment length, and potential data imbalance. Next, text preprocessing and tokenization are performed before dividing the data using stratified holdout splitting to preserve class proportions in training and testing sets. Both models are then trained with the same hyperparameter settings to ensure an objective and fair performance comparison. Results show that IndoBERT achieved an accuracy of 0.8333, while IndoBERTweet performed better with an accuracy of 0.8409. The analysis of the confusion matrix and ROC curve confirms that IndoBERTweet is more effective at detecting cyberbullying across different classes. Compared to previous studies using the SMOTE method and Bernoulli Naïve Bayes algorithm, which achieved 84.00% accuracy, this study's findings are slightly higher at 84.09%. Notably, this was achieved without using synthetic oversampling techniques. This suggests that the approach employed in this research can deliver competitive performance even without data balancing with SMOTE. Overall, these findings indicate that a transformer-based approach, combined with a more representative dataset, can improve cyberbullying detection more efficiently and practically. Therefore, IndoBERTweet is a more suitable model for implementing a cyberbullying content moderation system in Indonesia.]]>
