Optimasi Klasifikasi Sentimen Ulasan Game Berbahasa Indonesia: IndoBERT dan SMOTE untuk Menangani Ketidakseimbangan Kelas

Abstract

The increased use of gaming apps on platforms like the Google Play Store has signaled the importance of user reviews as a source of app quality evaluation. However, sentiment analysis of Indonesian-language reviews faces challenges due to the peculiarities of language structure, emotional expressions, and the use of slang and specialized terms in game reviews. This study aims to classify reviews into three sentiment classes: positive, negative, and neutral, using the IndoBERT-base-uncased model. The type of research used is experimental by comparing the performance of the model using original and synthetic datasets. The total original dataset collected was 998 reviews. The k_neighbors SMOTE parameter used is 5. The IndoBERT-base-uncased epoch parameter is 10, with a batch value per device and a batch for evaluation of 16. Configuration variable warmup_steps is 500 with L2 weight_decay regularization at 0.01. Evaluation results after SMOTE implementation: the precision score increased from 0.44 to 0.45, and the F1-score from 0.46 to 0.47. However, the recall score did not increase. The evaluation results show that the model has variable performance between classes with an initial accuracy of 69.,5%. Data imbalance is a major challenge, especially in minority classes such as class 1 (neutral), which cannot be predicted by the model. The SMOTE technique successfully improved data balance and increased accuracy to 72.5%, as well as improving metrics such as precision, recall, and F1-score overall.