Skripsi
PERBANDINGAN RANDOM MAJORITY UNDERSAMPLING DAN RANDOM OVERSAMPLING PADA ANALISIS SENTIMEN MENGGUNAKAN BERT UNTUK ULASAN ROBLOX
Game application reviews, such as those for Roblox, often exhibit significant class imbalance, where positively labeled reviews vastly outnumber negative ones. This study aims to compare the effectiveness of Random Majority Under-sampling (RMU) and Random Over-sampling (ROS) data balancing techniques on the Bidirectional Encoder Representations from Transformers (BERT) architecture for automatic sentiment classification. The initial dataset was sourced from 435,826 Roblox user reviews, which were subsequently extracted to form a final sample of 11,970 positive and 2,961 negative reviews. The test results from nine experimental scenarios demonstrate that the application of RMU triggers the loss of linguistic information and transforms the model into a single-class predictor (F1-score 0.444981). Conversely, the best configuration was achieved by BERT+ROS (Scenario 9, learning rate 5e-05), which successfully recorded an accuracy of 0.959812, precision of 0.926702, recall of 0.952050, and F1-score of 0.938640. Based on these findings, it is concluded that the BERT architecture optimized with the ROS technique is the most superior approach for imbalanced text-based sentiment analysis.