Cell phones show human movement predictable 93% of the time

[Eugen Leitl]

http://arstechnica.com/science/news/2010/02/cell-phones-show-human-movement-predictable-93-of-the-time.ars

Cell phones show human movement predictable 93% of the time

By Casey Johnston | Last updated February 23, 2010 5:02 PM

We'd like to think of ourselves as dynamic, unpredictable individuals, but
according to new research, that's not the case at all. In a study published
in last week's Science, researchers looked at customer location data culled
from cellular service providers. By looking at how customers moved around,
the authors of the study found that it may be possible to predict human
movement patterns and location up to 93 percent of the time. These findings
may be useful in multiple fields, including city planning, mobile
communication resource management, and anticipating the spread of viruses.

It's not currently possible to know exactly where everyone is all the time,
but cell phones can provide a pretty good approximation. Cell phone companies
store records of customers' locations based on when the customers' phones
connect to towers during calls. Researchers realized that taking this data
and paring it down to users who place calls more frequently might allow them
to see if they could develop any measure of how predictable human movements
and locations are. The users they worked with placed calls an average of once
every two hours, connecting to towers that cover an area of about two square
miles.

The authors analyzed various aspects of the information related to the calls,
as well as information that could be aggregated over multiple calls: number
of distinct locations, historical probability that the location had been
visited in the past, time spent at each tower, the order in which customers
usually visited towers, and so on. With these numbers, the authors could
create measures of the entropy of the customers' trajectories. To control for
uncertainty, they also looked at instances where a customer was not in
communication with the grid and effectively invisible to them, and removed
those that had frequent extended periods of invisibility.

Most customers seemed to stick to the same small area, a radius of six miles
or less, but there were a few callers that regularly traveled areas of a
radius of hundreds of miles. It would seem that the cell phone users who
traveled the least would be the most predictable in their movements, but the
authors found this to be untrue. All users were roughly equally predictable,
regardless of the size of their typical traveled region. Everyone seemed to
have a set area that they rarely left, and that area was always traveled in a
very regular waybeven the jet-setters appear to rarely deviate from their
travel patterns.

Customers that stuck to the same six-mile radius had predictability rates of
97 to 93 percent, and this fell off as the typical area of travel grew. But
the predictability eventually stabilized, and remained at 93 percent even as
the radius of travel rose to thousands of miles. Regardless of how widely
they traveled, the researchers could adequately predict their locations, down
to the specific tower, 93 percent of the time.

Breaking down the schedules of users by the hour allowed the authors to see
how the variability changed during the course of a day. As might be expected,
users' locations had the lowest measures of regularity during transition
periods, such as the hours before and after work and during lunch times.
Customers also had a 70 percent likelihood of being at their number one
most-visited location at any random point in time. That's quite a high
number, considering that randomizing positions over the average number of
locations visited per person gives a 1.6 percent likelihood of finding them
at each one.

The authors note that this research has a variety of practical implications.
Knowing how easy it is to predict human movement, mobile communications
businesses could anticipate data load (we're looking at you, AT&T) and city
planners could use the data to inform their models of traffic flow. The big
limitation of the study was the restriction of the analysis to fairly
frequent cell phone users, but it might be possible to combine this with
other data sets to form harder and faster human location predictions.

Science, 2010. DOI: 10.1126/science.1177170