Mining place-time affinity to improve POI recommendation

Abstract

A Point of Interest(POI) is a location that one may find useful or interesting. POI recommendation is a key feature in location-based social networks (LBSNs). With the development of mobile devices and apps, POI recommendation becomes a very popular topic and it includes humongous data. Current models always suffer from the problem of data sparsity. In this paper we propose a novel transfer learning model to learn affinity between the time and places, and use the mined features to improve the performance of a content-based POI recommendation system. In particular, we use check-in data to learn latent vectors for time and place category features by non-negative matrix factorization. Then, the mined densely embedded features are input to a gradient boosting decision tree (GBDT) based pairwise scoring model, which is trained by the check-in data of another city, to do POI recommendation. We conduct our experiment on the Foursquare check-in dataset, and discover that the learned latent vectors can dramatically improve the performance of a POI recommendation system.