Targeted service improvement for individualized providers

Online reviews offer valuable insights for service improvement, especially with the increasing availability of multimodal data (e.g., text and images). However, most existing research has primarily focused on industry- or group-level analyses, often overlooking individual business contexts (e.g., performance and competition), the interaction effects between different service attributes, and the full potential of image information. This study fills these gaps by developing a novel multimodal framework that leverages machine learning and optimization theory to generate targeted service improvement recommendations for individual service providers. The three-stage framework begins with extracting and clustering aspect-opinion pairs from review texts and matching them with corresponding image features. Subsequently, we use a feature distribution smoothing-based Bayesian iterative algorithm to address biases from data imbalance and estimate attribute-level interaction effects. Finally, we construct an interaction effect network, propose a probability-based stepwise optimization algorithm, and integrate image features with large language models to generate specific and actionable recommendations for individual providers. Through a large-scale case study of hotel service improvement, we validate the effectiveness of the proposed framework and demonstrate enhanced predictive accuracy and management insights. A user study with practitioners showed that our method outperformed the best baseline by 19.5 % in practitioner satisfaction.