<![CDATA[Recurrent Neural Networks (RNNs) such as LSTM and GRU are designed for sequential data. However, their performance in emotion detection is often compromised by class imbalance. This study compares LSTM and GRU architectures for classifying emotional states using a dataset of 4,386 Indonesian tweets. The dataset exhibits a mild imbalance (approximately 1.7:1) across five classes: Anger, Happy, Sadness, Love, and Fear. However, the effectiveness of these models is often hindered by class imbalance in datasets, which biases predictions toward majority classes and compromises the reliability of standard metrics. This study aims to systematically evaluate the comparison of LSTM and GRU architectures in processing imbalanced Indonesian emotional tweet data. The methodology involves evaluating these models across various resampling techniques, including Random Oversampling, SMOTE, and Near-Miss. Key findings reveal that LSTM consistently outperforms GRU in capturing complex emotional patterns. Specifically, the LSTM model combined with Random Oversampling emerged as the most robust configuration, achieving a Macro-F1 score of 71% and an accuracy of 73%. While Random Oversampling effectively enhanced minority class recognition without overfitting, SMOTE and Near-Miss introduced significant performance trade-offs. These results provide actionable insights for selecting optimal architectures and resampling strategies to mitigate imbalance-related biases in sequential classification tasks.]]>
Copyrights © 2026
