<![CDATA[Cyberbullying on social media platform X has become a critical digital threat and requires automatic detection mechanisms to mitigate psychological impacts on victims. This study proposes a hybrid deep learning architecture that combines Convolutional Neural Network (CNN) for local feature extraction and Long Short-Term Memory (LSTM) for sequential context understanding in classifying Indonesian language cyberbullying comments. This study evaluates model performance using a dataset of 13,677 comments from social media X through a series of systematic testing scenarios, including the impact of regularization, utilization of FastText embeddings, and comparative studies against state-of-the-art models. Experimental results demonstrate that the Early Stopping mechanism is a critical factor in this architecture, where without this mechanism the model experiences accuracy degradation of up to 32%. The proposed CNN-LSTM model achieves 88.38% accuracy and 88.00% F1-Score, improving to 0.9559 AUC with FastText integration. This model achieves over 97% of IndoBERTweet's performance with 22 times lower computational complexity (4.97 million versus 110.88 million parameters) and outperforms machine learning methods such as SVM with an accuracy margin of more than 10 percentage points. This study concludes that the CNN-LSTM architecture offers a robust and efficient solution for cyberbullying detection, particularly for resource-constrained environments]]>
Copyrights © 2026
