The role of context fusion on accuracy, beyond-accuracy, and fairness of point-of-interest recommendation systems

Point-of-interest (POI) recommendation is an essential service to location-based social networks (LBSNs), benefiting both users providing them the chance to explore new locations and businesses by discovering new potential customers. These systems learn the preferences of users and their mobility patterns to generate relevant POI recommendations. Previous studies have shown that incorporating contextual information such as geographical, temporal, social, and categorical substantially improves the quality of POI recommendations. However, fewer works have studied in-depth the multi-aspect benefits of context fusion on POI recommendation, in particular on beyond-accuracy, fairness, and interpretability of recommendations. In this work, we propose a linear regression-based fusion of POI contexts that effectively finds the best combination of contexts for each (i) user, or (ii) group of users from their historical interactions. The results of large-scale experiments on two popular datasets Gowalla and Yelp reveal several interesting findings. First, the proposed approach does not present significant loss in accuracy and unfairness of popularity bias as with classical collaborative baselines, and yet improves the beyond-accuracy of recommendation compared with existing context-aware (CA) approaches using heuristic context fusions; for instance, the proposed approach improves the accuracy and beyond-accuracy compare to best baseline model by 25% and 30%, respectively. Second, our proposed approach is interpretable, allowing to explain to the user why she has been recommended specific POIs, based on the learned context weights from user past check-ins; for example, if you are in Rome and our method recommends you a historical place like ‘Colosseum’, it can also provide an explanation why this item is recommended to you based on your personal preference on context (e.g., you were recommended to visit ‘Colosseum’ because in the past your visited historical places). Third, by analyzing the fairness of recommendation with respect to users (based on their activity levels) and items (based on the popularity of items), we found that a model which is recommend fairly on one dataset can recommend unfair on another dataset. Overall, our study suggests that appropriate context fusion is an essential element of an accurate, fair, and transparent POI recommendation system. We highlight that while we have tested the efficacy of our context-fusion methods on two popular CA recommendation models in the POI domain, namely GeoSoCa and LORE, our system can be flexibly utilized to extend the capability of other CA algorithms.