Context-aware personalized path inference from large-scale GPS snippets

Abstract Path inference is essential for many location-based services, and it infers vehicle moving paths from noisy and incomplete data captured by GPS devices mounted on vehicles. However, Path inference remains a highly challenging task due to uncertain measurements, dynamic road context and diverse personal preferences of drivers. Existing works usually mapped GPS observations to candidate paths by assuming the same road contexts and the same preference for all drivers, which is unreasonable in practice. To address those issues, we propose a graphical model ConPPI based on conditional random field for path inference, by simultaneously considering both the road average speed which is one of the most important contexts, and the personalized information in a unified model. To exploit these information, ConPPI learns the road average speeds from pair-wise location data and discovers the drivers’ personal preferences collaboratively. Extensive experiments on subsets from a real world dataset show that ConPPI can accurately infer both road average speeds and personal preferences, which leads to better path inference performance when compared with state-of-the-arts techniques.