<![CDATA[The increasingly massive phenomenon of bullying spreading on social media drives the urgency of applying computational approaches to detect cyberbullying quickly and accurately. This study aims to compare the effectiveness of the Support Vector Machine (SVM) algorithm based on TF-IDF against the Long Short-Term Memory (LSTM) architecture based on Word Embedding in classifying bullying texts. The experiment was conducted utilizing a dataset of 6,520 Indonesian comments extracted from the social media platform X (Twitter). The entire data was then divided into 80% training data and 20% testing data. Performance measurements proved that the SVM model dominated the overall prediction accuracy, achieving an Accuracy of 0.8352 (83%), Precision of 0.8356 (83%), and F1-Score of 0.8363 (83%). Conversely, the LSTM model demonstrated superiority in the sensitivity of positive class identification, scoring a Recall of 0.8552 (85%) and ROC-AUC of 0.9088 (90%). In terms of computational efficiency, the SVM model proved to be far more resource-efficient, requiring only 42.90 seconds of training time and 74.87 MB of memory, in contrast to the LSTM which consumed up to 106.67 seconds and 126.80 MB of memory. This research concludes that SVM is the most optimal method overall because it is able to balance high detection accuracy with excellent operational load efficiency.]]>
Copyrights © 2026
