own experiments to determine how motion was reflected in CDRs. We ignored text CDRs because text messages involve only a single location. Second, since we were interested in routes to and from the center of town, we used only CDRs with antenna sequences that began or ended at the tower handling calls for the core downtown area. After filtering, we were still left with tens of thousands of CDRs.

We began by identifying 15 common commuting routes ( 13 driving routes and two train routes) radiating from the town center. We obtained ground-truth data for them by driving/ riding each one four times (two in each direction), using at least two phones calling each other on each drive/ride. We obtained the CDRs for these calls to both train and test our algorithms. From our training data, we determined a reference pattern of cellular sectors used by calls on each of the routes. We intentionally included some routes very close to one another and others that partially overlap, as routes do in real life. Some of our reference patterns were thus quite similar, making disambiguation a challenge.

We then developed two methods for assigning CDRs to routes: One uses a distance metric to assign a test CDR to the route with the closest reference pattern. We used a variant of Earth Mover’s Distance (EMD), a measure of the difference between two arbitrary probability distributions, as a metric that takes into account common subsets of sectors, the particular sequence of sectors, how long the call is associated with each sector, and tower locations. The other method uses as reference data the radio-fre-quency scans routinely performed by cellular network operators to measure network coverage. The scanner data contains signal-strength measurements stamped with global-position-ing system (GPS) locations from all observable antennas along major driving routes. Our classification algorithm estimates the likelihood of a given sequence of antennas being seen on a particular route and selects the most likely route. This approach has the advantage of being able to reuse data that is already available, without requiring additional data collection on every target route. It could easily be ex-

figure 4. Paradeshed of Morristown, NJ; the red dot denotes the city center.

 

Five times as many people were in Morristown for the st. Patrick’s day Parade as on a normal saturday. to show the geographical distribution of parade attendees’ homes, we mapped the number of people coming from each surrounding ZIP code. Green-yellow areas contributed more than the parade-day average and purple-red areas less than that average. Communities contributing near the average are not colored to highlight the outliers.

 

of a geographic area during arbitrary time periods. Of particular interest to city officials is how the mix of inhabitants changes during special occasions (such as extreme weather, construction projects, and regional events). Knowing where people come from can help them in advertising for the event and easing traffic congestion.

One such occasion in Morristown was the St. Patrick’s Day Parade on Saturday March 12, 2011, from 11 a.m. to 3 p.m. We repeated our analysis for obtaining the laborshed but on cellphone transactions handled during the time of the parade by the antennas pointing along the parade route. Figure 4 is the resulting paradeshed, with people coming for the parade, compared with data for the same antennas and time interval on a typical Saturday without special events. The parade is a county affair, so we would have expected the event to draw widely from other parts of the county (north and west of Morristown). Indeed, we see the areas north and west of Morristown showing large increases, while other areas south and east show smaller increases. Prior to the

;

New York

10

New Jersey

1

3

6

 

advent of cellular networks, it was notably difficult for local officials to obtain this information except through expensive surveys.

traffic Volume

The quality of life in any urban area is directly influenced by the frustration, pollution, time lost, and noise of traffic congestion. Efforts by planners to improve traffic flow while not sacrificing street life need a thorough understanding of existing traffic conditions. Since traditional methods of obtaining traffic data are expensive, we set out to determine whether we could estimate traffic volumes from CDRs.

To explore traffic volume on major commuting routes into Morristown, we used the same data-collection procedure we used to calculate the laborshed, as described earlier. However, in this case we recorded activity in and around Morristown from December 2009 to January 2010. We used two filters to obtain an appropriate subset of CDRs for the study: First, to retain data about moving vehicles, we used only voice CDRs including antennas on at least five towers, as indicated by our