• data science
• feedback loop
• human mobility
• recommender systems

Recommender systems (RSs) have revolutionized the way humans interact with digital applications. From social networks (e.g. Facebook and X) to online selling platforms (e.g. Amazon and eBay), almost every system adopts a recommendation strategy to enhance the user experience.
In urban ecosystems, RSs engage with individuals by suggesting sites to visit (e.g., Foursquare and Tripadvisor) and routes to reach their destinations (e.g., Google Maps and TomTom), connecting with thousands of people in real time.
Research across different domains has revealed numerous unintended consequences associated with RSs applications, but their impact on the urban ecosystem remains largely unexplored. This thesis aims to develop an experimental environment to simulate the interactions between a next-location RS and a group of users. The objective is to observe how a continuous use of the RS influences mobility behaviours and affects the performance of the model over time. Three recommendation policies are selected and tested with the same pool of users to study different recommendation paradigms as the acceptance rate of the suggestions increases.
The results indicate that when RSs are retrained without proper oversight, recommendations become biased toward popular or previously visited locations, reducing diversity and marginalising less visited locations.
This bias worsens when AI-based RSs are employed. In POI similarity-based RSs, the same issue arises unless users consistently accept the suggestions. However, RSs that rely on user similarity help mitigate the bias, though they still lead to a consistent decline in user trajectory entropy.