Balancing happiness and efficiency in POI recommendation via Kahneman–Tversky optimization

Abstract The recommendation of points of interest (POIs) is essential for location-based services, particularly targeted advertising and content delivery. However, existing systems prioritize efficiency and recommend easily accessible locations based on past mobility data. A quantitative framework that simultaneously considers human emotional states, such as happiness and efficiency, has not been fully established. This study proposes a POI-recommendation model that aims to balance happiness and efficiency by applying the Kahneman–Tversky optimization (KTO), an optimization method based on prospect theory. Our approach first extracts individual POI sequences from GPS trajectory data and constructs a time-series generative model. We then fine-tune the model using KTO so that it can generate POI sequences that are both easy to visit and likely to increase personal happiness. Experimental results using large-scale mobility data show that the proposed model retains approximately 80% of the prediction performance for novel POIs, while incorporating mobility characteristics associated with higher happiness levels. The model recommends sequences with fewer routine POIs and a higher proportion of leisure or entertainment facilities. It also shows tendencies toward longer travel distances, greater diversity, and increased mobility-pattern irregularities. Furthermore, we demonstrate that adjusting user mobility preferences within the model can control these recommendation tendencies. These findings support the establishment of a methodological framework for POI recommendations that jointly prioritizes human emotions and efficiency, with happiness as a key example. The proposed framework can be widely applied to design user experiences that promote appropriate behavioral changes by leveraging the actual mobility behaviors and preferences of users from the perspectives of emotions and efficiency. Future research should present these recommendations to real users and analyze how user acceptance and self-reported happiness levels change accordingly.