Topic Analysis Using LDA-LSTM on Shopee User Comment

Shannon Dominique Saputra (Soegijapranata Catholic University)
Albertus Dwiyoga Widiantoro Dwiyoga Widiantoro (Soegijapranata Catholic University)

Sisforma: Journal of Information Systems
Vol 13, No 1: May 2026

Shopee's growth has also shaped online shopping in Indonesia. This study analyzed user reviews to measure satisfaction and identify key service issues using a hybrid framework: LDA for topic modeling and LSTM for sentiment classification. Class imbalance was addressed using a combination of Random Oversampling and Neighborhood Cleaning Rule (ROS-NCL).The results showed that LSTM + ROS-NCL outperformed ROS, NCL, and SMOTE, with 95% accuracy and a precision, recall, and F1-score of 0.94 each. These findings demonstrate that oversampling combined with noise cleaning effectively improves performance on imbalanced data, while also providing practical insights for feature development, logistics improvements, and Shopee's promotional strategy.