Learning to Rank for Spatiotemporal Search

In this article we consider the problem of mapping a noisy estimate of a user's current location to a semantically meaningful point of interest, such as a home, restaurant, or store. Despite the poor accuracy of GPS on current mobile devices and the relatively high density of places in urban areas, it is possible to predict a user's location with considerable precision by explicitly modeling both places and users and by combining a variety of signals about a user's current context. Places are often simply modeled as a single latitude and longitude when in fact they are complex entities existing in both space and time and shaped by the millions of people that interact with them. Similarly, models of users reveal complex but predictable patterns of mobility that can be exploited for this task. We propose a novel spatial search algorithm that infers a user's location by combining aggregate signals mined from billions of foursquare check-ins with real-time contextual information. We evaluate a variety of techniques and demonstrate that machine learning algorithms for ranking and spatiotemporal models of places and users offer significant improvement over common methods for location search based on distance and popularity.