Place Recommendation with Implicit Spatial Feedback

Since the advent of the Netflix Prize [1], there has been an influx of papers on recommender systems in machine learning literature. A popular framework to build such systems has been collobarative filtering (CF) [6]. On the Netflix dataset, CF algorithms were one of the few stand-alone methods shown to have superior performance. Recently, web services such as Foursquare and Facebook Places started to allow users to share their locations over social networks. This has led to an explosion in the number of virtual places that are available for checking in, inundating users with many irrelevant choices. In turn, there is a pressing need for algorithms that rank nearby places according to the user’s interest. In this paper, we tackle this problem by providing a machine learning perspective to personalized place recommendation. Our contributions are as follows: First, we transform and formalize the publicly available checkin information we scraped from Twitter and Foursquare as a collobarative filtering dataset. Second, we introduce an evaluation framework to compare algorithms. The framework takes into account the limitations of the mobile device interface (i.e. one can only display a few ranked places) and the spatial constraints (i.e. user is only interested in a ranking of nearby venues). Third, we introduce a novel algorithm that exploits implicit feedback provided by users and demonstrate that it outperforms state-of-the-art CF algorithms. Finally, we discuss and report preliminary results on extending our CF algorithm with explicitly computed user, place, and time features.