Interpretable machine learning for customer satisfaction optimization

Purpose Understanding how dining experiences are transformed into customer satisfaction judgments is central to hospitality competitiveness. Using high-dimensional user-generated content on tourism dining experiences, this study aims to examine whether shared three-factor classifications coexist with segment-specific differences in satisfaction relationships. Design/methodology/approach The authors analyze 82,105 Yelp restaurant reviews from nine market segments based on price level, cuisine type and restaurant category. Bidirectional encoder representations from transformers-based aspect mining and clustering extract attribute–sentiment pairs. Ensemble models estimate nonlinear predictive relationships, and SHAP decomposes their predictions to support attribute classification, segment-specific satisfaction functions and diagnostic prioritization. Findings Empirical results reveal a basic-dominant pattern together with segment-level variation in estimated penalty–reward magnitudes, satisfaction-function shapes and diagnostic priorities. This pattern remains largely stable under alternative impact asymmetry thresholds and the procedure based on penalty-reward contrast analysis. Categorical stability can therefore coexist with segment-contingent differences in evaluative weight and relationship form. Practical implications Theoretically, the findings highlight the value of extending attribute classification with segment-specific satisfaction-function analysis, because segment heterogeneity is expressed within, and not only between, three-factor categories. Practically, managers can use the framework to diagnose service attributes with higher marginal predictive utility and develop more targeted, segment-specific improvement priorities. Originality/value This study refines the categorical interpretation of three-factor theory by showing that a shared attribute classification does not imply an invariant satisfaction relationship. Segment heterogeneity can occur within a shared category through differences in penalty–reward magnitude and satisfaction-function shape, not only through category shifts.