Beyond Prediction: Using Explainable AI to Uncover the Drivers of Customer Ratings with SHAP and Review Topics

Abstract

Understanding the drivers of customer satisfaction is significant for luxury restaurants to maintain and increase consumers. While online reviews provide a rich source of customer opinions, extracting clear and actionable insights from this unstructured text carries a major difficulty for decision makers. This paper proposes a framework to not only predict customer star ratings from reviews, but also to explain the key drivers behind those predictions. BERTopic is applied on a dataset of reviews from Michelin-starred restaurants in Turkey. Four different feature engineering strategies are compared in this study, changing based on the integration of topic modeling and sentiment analysis. Five machine learning models were leveraged to predict the star rating, and the SHAP framework was used to interpret best-performing models and select best drivers of prediction. The results show that feature sets which include sentiment outperform those based on only topic presence. Weighted sentiments feature achieved the best predictive performance, with the XGBoost yielding the lowest error. The SHAP analysis revealed that the model’s predictions are mainly driven by primary performance factors like Dining Experience and Staff Service. Staff Service and Reservation Process were found as features for which a failure strongly lowers the rating, while a success provides almost no benefit on ratings. This research provides a contribution with successfully identifying and quantifying the key drivers of customer satisfaction. By combining topic modeling with explainable AI, this work moves beyond prediction to provide actionable insights.