Deep Learning-Based Analysis of Low-Rated Thai App Reviews on the Google Play Store

Abstract

This study aims to develop a Deep Learning model for classifying problem types from low-rated (1–3 stars) Thai-language reviews on the Google Play Store, which serves as an important data source reflecting real user experiences and application issues. These reviews are unstructured textual data and exist in large volumes, making manual analysis difficult and inefficient.

           The research formulates the task as a four-class multi-class classification problem, consisting of the following categories login, payment, performance, and other. The dataset comprises 1,998 labeled reviews. The XLM-RoBERTa model was fine-tuned using the Cross-Entropy Loss function and the AdamW optimizer.

           Experimental results on a test set of 400 reviews show that the model achieved an accuracy of 0.71 and a macro F1-score of 0.73. The payment class demonstrated the highest classification performance, whereas the other class exhibited partial overlap with some of the remaining classes.

           In addition, a prototype web application was developed to automatically retrieve reviews, classify problem types, and visualize the results through a dashboard to support decision-making processes. The findings indicate that Transformer-based techniques can be effectively applied to Thai-language review analysis and can efficiently transform unstructured textual data into structured information.