Point-of-interest lists and their potential in recommendation systems

Location based social networks, such as Foursquare and Yelp, have inspired the development of novel recommendation systems due to the massive volume and multiple types of data that their users generate on a daily basis. More recently, research studies have been focusing on utilizing structural data from these networks that relate the various entities, typically users and locations. In this work, we investigate the information contained in unique structural data of social networks, namely the lists or collections of items, and assess their potential in recommendation systems. Our hypothesis is that the information encoded in the lists can be utilized to estimate the similarities amongst POIs and, hence, these similarities can drive a personalized recommendation system or enhance the performance of an existing one. This is based on the fact that POI lists are user generated content and can be considered as collections of related POIs. Our method attempts to extract these relations and express the notion of similarity using graph theoretic, set theoretic and statistical measures. Our approach is applied on a Foursquare dataset of two popular destinations in northern Greece and is evaluated both via an offline experiment and against the opinions of local populace that we obtain via a user study. The results confirm the existence of rich similarity information within the lists and the effectiveness of our approach as a recommendation system.