Communications - January 2013

tended to larger-scale studies in other urban areas.

Both classification algorithms achieved approximately 90% accuracy on our test data, outperforming several other algorithms based purely on common subsets of towers, sectors, or antennas. Our route-classification algorithms and their accuracy are described in more detail in Becker et al. 1

Figure 5 shows the result of our route assignment to moving phones in the Morristown area, using the EMD-based algorithm applied to CDRs; the signal-strength-based method yields similar results. The widths of the lines superimposed on each route are proportional to the estimated traffic volumes on each route. The two wide black lines running roughly north and south correspond to the interstate highway that passes through Morristown. The counts shown at the beginning of each route are normalized to 1,000 moving phones. We compared our relative traffic volumes to traffic counts published by the New Jersey Department of Transportation17 and found a correlation coefficient of 0.77, giving us added confidence in the accuracy of our approach.

ine differences in behavior between tourists and locals in New York City. 9 Calabrese et al. 6 studied where people came from to attend special events in Boston, finding that people who live close to an event are more likely to attend it and that events of the same type attract people from roughly the same home locations. Though we have also studied how cellular network data can be used for urban planning, we pursued different research goals (such as calculating daily ranges, deriving and validating laborsheds, and estimating traffic volume).

In the domain of mobility model-
ing, Gonzalez et al. 10 used cellular net-
work data from an unnamed European
country to form statistical models of
how individuals move, finding human
trajectories reflect a high degree of spa-
tial and temporal regularity, with each
individual having a time-independent
characteristic travel distance and re-
turning often to a few characteristic
locations. Song et al. 21 analyzed similar
data to study the predictability of an in-
dividual’s movements, finding a high
degree of predictability across a large
user base largely independent of travel
distances and other factors. Whereas
these efforts modeled individuals, we
focused on mobility differences be-
tween large populations in distinct
geographic regions.

figure 5. Relative traffic volume on 12 commuting routes to the center of Morristown, NJ, as assigned by our route-classification algorithms.

Related Work

The research community increasingly uses cellular network data to study human mobility, applying its findings to various domains, including urban planning, 19 mobility modeling, 10 so-cial-relation inference, 11 and health care. 3 Here, we survey a subset of that work most similar to our own.

Several efforts have explored how cellular network data can be used for urban planning. In studies of Milan, Italy, Ratti et al. 19 and later Pulselli et al. 18 demonstrated it is possible to characterize the intensity and spatiotemporal evolution of urban activities using call volume at cell towers. Reades et al. 20 studied call-volume activity in six distinct locations in Rome, Italy, showing that volume varied drastically between the studied locations and between weekdays and weekends. Girardin et al. 8 used tagged photographs from Flickr in combination with call-volume data to determine the whereabouts of locals and tourists in Rome. They later repeated the study with only call-volume data to exam-