The increasing use of the internet in Indonesia has an influence on the presence of Online Travel Agents (OTA). Through the OTA application, users can book transportation and accommodation tickets more easily and quickly. The increasingly rigorous competition is causing companies like PT XYZ to be able to provide solutions to the needs and problems of their customers in the field of online ticket booking. Many customers submit reviews of the use of the PT XYZ application through Playstore and Appstore, and it needs a technique to group thousands of reviews and detect the topics discussed by customers automatically. In this study, we classified reviews from Android and iOS applications using BERT that had been adjusted through fine-tuning with IndoBERT, as well as modeling topics using LDA to evaluate the coherence score of each sentiment. The result of the comparison of hyperparameter models for the most optimal classification is epoch 4 with a learning rate of 5e-5. The accuracy obtained is 0.91, with an f1-score of 0.74. In addition, testing was carried out to compare BERT with other traditional machine learning. The best performing algorithm was Logistic Regression using TF-IDF word embeddings, achieving an accuracy of 0.890 and an F1-score of 0.865. Therefore, it can be inferred that the accuracy achieved by the fine-tuned classification model of IndoBert is sufficiently high for application in the PT XYZ review classification. Using a coherence score, we found 29 positive topics, 6 neutral topics, and 3 negative topics that were considered the most optimal. This finding can be used as evaluation material for PT XYZ to provide the best service to customers.
31 March 2024 in Jurnal Nasional Pendidikan Teknik Informatika (JANAPATI)
Forecasting the stock market is one of the most popular topics to be discussed in many fields. Many studies, especially in information technology have been conducted machine learning algorithms to achieve a more accurate prediction of the stock market. This research aims to find the effectiveness in predicting stock market performance by utilizing social media sentiment in combination with historical data. In addition, this research uses a machine learning algorithm to train a model to predict the stock price of each bank and training the model on a dataset that included the historical stock prices of the bank, as well as the sentiment scores of the social media posts about the bank and evaluate the performance of the model by comparing the predicted stock prices to the actual stock prices. The research shows that the R2 and RMSE score model that has been built with its historical data has slightly better performance than the model that has been built with the combination of historical data and social media sentiment. The finding indicates that the research method is closely correlated and affected to the performance of the stock market prediction.