Probabilistic Inference of Unknown Locations: Exploiting Collective Behavior when Individual Data is Scarce

In recent years there has been a proliferation in the use of large-scale, passively collected digital trace data to study the mobility and migration patterns of individuals in developing countries. Analysis of mobile phone and social media data, among other sources, has immediate policy applications that range from disease monitoring and city planning to disaster management and humanitarian relief. Unfortunately, existing methods for mining location-based information from passively collected data are generally not well suited to a large number of individuals in developing countries. This is in part due to the fact that technology use is quite heterogeneous, and that the lower intensity use patterns of many individuals produces a sparser digital trace.

In this paper, we present a method for predicting the approximate location of a mobile phone subscriber that is more appropriate to contexts where the signal generated by each individual may be intermittent, but the collective population generates a large amount of data. This method works well when, for instance, an individual is not consistently active on the network or when the phone is off. Our model uses a nonparametric approach to probabilistically interpolate locations, and has the advantage of associating a confidence with each prediction. We test this method on a large dataset of anonymized mobile phone records from Afghanistan, and find that we can correctly predict a subscriber's unknown location in 76%-95% of cases, and that on average our predicted location is off by 0.2-1.9 kilometers.