A RECURRENT NEURAL NETWORK–BASED SENTIMENT ANALYSIS OF MOBILE LEGENDS APP REVIEWS

Abstract

With the rapid growth of mobile applications, user reviews have become a valuable source of feedback for developers. This study investigates the use of a Recurrent Neural Network (RNN) for sentiment analysis of Mobile Legends user reviews. The textual data were preprocessed through cleaning, tokenization, and padding, while sentiment scores were converted into categorical labels. A Sequential RNN model, consisting of an Embedding layer, a SimpleRNN layer, and a Dense output layer with softmax activation, was constructed to classify reviews into three sentiment categories: negative, neutral, and positive. During training, the model achieved approximately 75% accuracy, and the Mean Squared Error (MSE) was 0.1354. However, evaluation using the classification report and confusion matrix revealed that the model was biased toward the negative class due to class imbalance, failing to correctly classify neutral and positive reviews. The high overall accuracy was misleading, as the model’s performance was limited by the dominance of the negative class. These results highlight the limitations of using a simple RNN architecture for multi-class sentiment classification in imbalanced datasets. To improve performance, future work should consider balancing the dataset through resampling or synthetic data generation and employing more advanced sequential models, such as LSTM or GRU, possibly combined with attention mechanisms or pretrained language models, to better capture the characteristics of all sentiment classes.