From destination image to regional embeddings: a multi-scale framework for cross-city tourism recommendation

Regional heterogeneity poses a persistent challenge for cross-city tourism recommendation, as tourist behaviour and destination characteristics vary widely across cultural, economic, and spatial contexts. This study investigates how such regional differences shape visitor behaviour and proposes a novel recommendation framework, FPNRegionalRec (Feature Pyramid Network-based Regional Recommendation Framework), that explicitly models city-level contextual factors. Using large-scale user-generated reviews collected from Ctrip across six Chinese cities (Beijing, Shanghai, Wuhan, Nanjing, Hangzhou, and Suzhou), we analyze cross-city differences in experiential, temporal, and geographic patterns and incorporate them into an interpretable recommendation model. Experimental results show that FPNRegionalRec consistently outperforms mainstream baselines, including collaborative filtering, graph-based, and sequential recommendation models, particularly in culturally similar and high-traffic cities, while regional embeddings remain robust under cold-start conditions. By connecting destination image theory and tourist behaviour heterogeneity with computational modelling, this study provides an interpretable framework for cross-city tourism recommendation. The findings can help online travel platforms and destination management organisations adapt recommendation strategies across heterogeneous cities, identify context-specific behavioural signals, and support more targeted point of interest (POI) promotion and cold-start management.