Consider Figure 1a, where two cars run on parallel streets before turning
in opposite directions. Figure 1b shows how the cars are observed with GPS
equipment, every dot is an observation. Given the current observations and an
history of previous trips of the user, a system can match the current trip with
the repository to nd the most probable destination for it. This is possible when
there is a sucient number of previous trips and one of them matches to some
degree with the actual one. But what happens with noisy observations, such as in
Figure 1c? Inferring location data with cell-towers is a dicult task: real-world
data shows that the approximation in urban areas varies between 400 and 1600
meters. Predictor systems have problems in the matching step when locations
are inferred with cell-towers, thus the prediction can fail.