Do We Need Real Data? - Testing and Training Algorithms with Artificial Geolocation Data

Abstract

As big data becomes increasingly important, so do algorithms that operate on geolocation data. Privacy requirements and the cost of collecting large sets of geolocation data, however, make it difficult to test those algorithms with real data. Artificially generated data sets therefore present an appealing alternative. This paper explores the use of two types of neural networks as generators of geolocation data and introduces a method based on the Turing Test to determine whether generated geolocation data is indistinguishable from real data. In an extensive evaluation we apply the method to data generated by our own implementation of neural networks as well as the widely used BerlinMOD generator on the one hand, the four most prominent data sets of real geolocation data covering at total of 65 million records on the other hand. The experiments show that in eleven of twelve cases artificial data sets can be told from real ones. We conclude that, at present, the generators we tested provide no safe replacement for real data.