Estimating Online User Location Distribution without GPS Location

Abstract

We focus on the problem of offline user location estimation using online information, particularly for the application of TV segment advertising. Unlike previous works, the proposed method does not assume GPS information, but works with loosely structured information such as English location description. We propose to use a neural language model to capture the semantic similarity among the location descriptions. The language model can help reduce the otherwise expensive geolocating service lookups by internally resolving similar areas, neighborhoods, etc. onto the same description. We also propose a metric for comparing geodemographic histograms. This metric considers the demographic gap between the online world and the offline world. In the experiments section, we demonstrate the recall and accuracy of our language-based, GPS-free user location distribution estimation. In addition, we illustrate the effectiveness of the proposed distribution estimation metric.