This paper (https://lnkd.in/gtZFi4_d) suggests an interesting approach. Keeping this specific use case aside, the generalized methodology here, with a few different variations, can help you not only transform recommendations on your online channels, but can also help enhance your “nudging theory” based tools.

This study addresses the challenge of personalized tourism recommendation in the domain of cultural heritage sites: recommending not just places but routes, experiences, and balancing visitor preferences with preservation constraints.

They propose a deep-learning framework composed of three modular innovations that specifically target the complexities of heritage tourism:

1. Heritage-aware Graph Neural Networks (H-GNNs): to model propagation of “cultural significance” between heritage sites and capture temporal/visitor‐behavior dynamics.

2. Cultural Spatiotemporal Transformers (C-STTs): to optimize visitor routing (both spatial and temporal) while respecting heritage-site constraints (opening hours, capacity, preservation priorities).

3. Heritage-adaptive Transfer Reinforcement Learning (H-TRL): to adapt recommendation policies across different heritage contexts, with reinforcement‐learning style reward structures that embed preservation/visitor‐satisfaction trade-offs.

The authors evaluate on several tourist/trajectory datasets (e.g., check-in data from Foursquare, Gowalla; GPS trajectory datasets) and demonstrate measurable improvements in behavior prediction, route optimization, and relevance of recommendations.

Going through the methodology, you can easily see the wide gamut of applications beyond this use case. An interesting approach indeed!