<![CDATA[The rapid growth of the online gaming industry in Indonesia has prompted developers to address various challenges in creating successful mobile games. This study aims to evaluate the effectiveness of ensemble learning techniques, particularly soft voting, in enhancing sentiment analysis accuracy across 17 genres of mobile games. Additionally, it identifies the most effective deep learning model for sentiment classification. The research compares the performance of CNN-LSTM, BERT, and CNN-GRU models, as well as an ensemble of these models. Review data was collected from the Google Play Store, then labeled and cleaned to improve data quality, categorized into positive, neutral, and negative sentiments. Data preprocessing techniques included cleaning, case folding, tokenization, normalization, stopword removal, and stemming. Word embedding techniques used were Word2vec for CNN-LSTM and CNN-GRU models, and IndoBERT for BERT model. Ensemble learning combined predictions from these models, significantly improving classification accuracy. Results indicate IndoBERT achieved an accuracy of 89%, while CNN-GRU and CNN-LSTM showed accuracies around 84-85%. The ensemble approach using soft voting successfully increased overall accuracy to 90% by combining predictions from all three models. The study concludes that ensemble learning effectively combines individual model strengths to enhance sentiment classification accuracy. Furthermore, user preference visualization for game genres revealed high popularity for "Strategy", "Word", and "Trivia" genres, while "Sports" genres were less favored. This research is expected to contribute to game developers in determining which genres to develop to enhance success chances and user satisfaction.]]>
