Explainability of Review Rating Prediction for M-Government Apps in Saudi Arabia

Mobile app reviews provide valuable insights into user needs. Predicting app ratings helps decision-makers respond to these needs. Natural Language Processing (NLP) has seen significant advancements, particularly with transformer-based models. However, these models are often considered “black box” and lack the ability to explain their predictions, making it difficult to trust their results. To address this challenge, this paper proposes a rating prediction explainability framework. It applies various transformer-based models to predict the ratings of more than 10,000 mobile government (m-government) app reviews developed in Saudi Arabia in English and Arabic and explains these predictions using SHAP and LIME explainability models. This work aims to improve our understanding of user needs from the m-government apps by analyzing their reviews and providing clear explanations for predictions, ultimately contributing to the improvement of app development processes. The transformer-based model ELECTRA achieved 96% accuracy with the English dataset while AraBERTv2 achieved 95% accuracy with the Arabic dataset.