Quantifying and mapping uncertainty in urban sentiment prediction: a combined approach with entropy and SHAP explanations.

Introduction: Mental health issues have been increasing globally, possibly linked to urbaniza¬tion and related lifestyles. There's a growing awareness that different aspects of the urban environment can influence mental health, either enabling or restrict¬ing behaviors that affect well-being. Recent studies have increasingly used social media data and machine learning techniques to assess urban well-being. However, these approaches often lack interpretability and do not explicitly account for prediction uncertainty, limiting their reliability for spatial decision-making. Methods: This study uses sentiment labels derived from georeferenced posts on X (formerly Twitter) via the NRC Emotion Lexicon. Supervised machine learning models are trained to predict sentiment based on urban environmental variables and generate spatial predictions for Lisbon, with performance compared across K-Nearest Neighbour (KNN), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Neural Network (NN). To go beyond simple point predictions, uncertainty at each location is measured with Shannon Entropy, based on class probabilities. Post-hoc explanations using SHapley Additive exPlanations (SHAP) identify urban features most impacting sentiment predictions. Results and discussion: Uncertainty maps combined with SHAP outputs show where interpretations are more reliable (low uncertainty) and where caution is needed (high uncertainty). In Lisbon 2019, SHAP identified proximity to cycling networks, fitness facilities, and NDVI as key factors influencing sentiment. This framework enhances spatial understanding of sentiment-environment relationships and offers a transparent way to identify areas where predictions are more reliable.