POI recommendation system : Formalization and evaluation

Abstract

The rise of location-based social networks (LBSNs) such as Foursquare, Facebook, and Geolife has revolutionized user interactions by capturing rich data such as check-ins, preferences, and movement patterns. In this context, point-of-interest (POI) recommendation systems (RS) play a crucial role in guiding users to relevant locations. However, traditional approaches, particularly collaborative filtering, have limitations in addressing the complexity of spatial, social, and temporal user behaviors. This thesis primarily addresses the inability of classical similarity measures to capture contextual proximity. Three major contributions are proposed: (1) the SPPUR model, which introduces a novel similarity measure inspired by the TF-IDF method, combining trajectory sequences with the geographic proximity between users, (2) the IPUMC model, which integrates implicit check-in similarity with an explicit measure of geographic distance, and (3) the IUPJS model, based on the Jaccard index and enhanced with a spatial component derived from users' start and end points. Empirical evaluations on Foursquare datasets (New York and Tokyo) confirm the superiority of these three models over existing approaches and highlight the importance of integrating contextual factors such as location, visit order, and social relationships into POI recommendation systems