Avaliação de embeddings geográficos em sistemas de recomendação de POIs.

Abstract

The Points of Interest (POI) recommendations are gaining prominence in the context of recommender systems (RSs), especially with the growth of Location-Based Social Networks (RSs), such as Foursquare, Gowalla and Yelp. The quality of these recommendations is essential to enrich the user experience on these platforms, facilitating sociability and promoting tourism, in addition to raising a series of challenges for the community. However, traditional POI recommendation systems are often limited to considering information such as location reviews, photos, access times and check-ins, neglecting relevant geographic data, such as geographic features that include rivers, buildings, streets and lakes in the context of a POI. These features can significantly influence user preferences, since users may visit a POI because they like the geographic features present in the environment. For example, some people prefer coffee shops near lakes and wooded areas rather than busy highways. In this study, we propose and evaluate the use of geographic embeddings that incorporate geographic features to improve POI SRs. The results indicated that the use of embeddings that consider the features increased the accuracy and MRR by up to 1.65% in the next POI recommendation task in the dataset used, compared to the baseline, confirming the importance of geographic features to improve POI SRs.