Inference in Complex Social Systems (Nathan Eagle) – ETech08

“Insights and Applications from the Behavior of the Aggregate”. Using cell phones as trackable tags, then extrapolating patterns. Learning about the aggregate by sampling the individual.

Nathan is a research scientist at MIT & Santa Fe. Also holds positions across sub-Saharan Africa.

Mobile phones are the fastest tech adoption in human history. People have extraordinary processing power and access to data. A new era of wearable computing.

Data, Science, and Engineering: Social network analysis. Classical social net metrics breakdown quickly as networks grow.

Demo: dynamic display of individual people walking around Cambridge, Mass, making cell calls. After accumulating data over time, can we make predictions about behaviors? What happens when we extend this to dyads of people? What relationships can be inferred from where and when dyads are? What about aggregate behavior of the whole? Do patterns emerge based on outlying events?

Data being logged in this trial: celltower ID’s, proximate bluetooth device name/activity, phone call/text log. Obviously huge privacy implications. All subjects were informed of logging. Have accumulated over 400,000hrs continuous human behavior data collected over 2004-2005.

Transitional probabilities used to evaluate eg likelihood of being at home versus being at work. Information entropy = ratio of amount of structure to randomness in subject’s routine. Shows variations between highly habitual individs and more random people. Low entropy subject vs. high entropy subject (which one are you?). “The Entropy of Life”. This data can be mapped against demographics to see what lifestyles are more or less entropic.

These models can be extended to map and model infectious patterns of contagions through social nets. Higher entropy individuals make containment much more difficult. (Work in progress).

Eigenbehaviors: A way to reduce highly-dimensional behavior data into a set of vectors that characterize individual behavior, but also behavior of demographics. Can a subject’s affiliation (demographic) be inferred from behavior patterns? Behavior space allows inference of demographic with high accuracy (90%).

Friendship vs. proximity Networks: can friendship be inferred from proximity? Behavioral signatures – friend vs acquiantance. Properties: Prox on Saturday night, phone communication, number of unique locs, prox outside work, prox at work, prox at home.

These models allow inferences about the true topology of social/friend networks. Data from mobile phones allows a much richer picture of social nets.

Organizational Rhythms: how the deadlines of an institution can be seen in the collective behavior of its individual members.

Network data mining: scale to 250 million nodes (phone #’s). Telecomm corps are very interested. 5,000 calls/sec; 12bil calls/month. Anonymized. Highly statistical averages are yielded every day. Furthermore, monthly plots are highly consitent from month to month. Ie human behavior across large numbers is highly organized over time. Why does the symettry exist? Why is the monthly curve of cell use the same every month for millions? What patterns/events coordinate or influence this behavior?

Diversity of your social net seems to correlate with positive socio-economic accomplishment (ie mo’ money & success).

Life inferences: Sleeping, Lunch = easy. Partying? Trickier. Auto diary to track your behaviors. How much sleep did I get? What did I do last Saturday night after midnight? How much time do I spend driving? Can I make predictions about my life?

The importance of triangles and mutual influence. Product adoption can be correlated in friend triangles. Friends have more influence over your purchasing.

Eprom – Educating sub-Saharan students on use of mobile data mining. SMS bootcamp. Mobile programming. Making epidemiology inferences from behavioral patterns, eg malaria susceptibility. Reality mining Africa. SMS Bloodbank, BoonaNet.

Takeaways:
Individual behavior prediction; relationship inference; organizational rhythms & aggregate behaviors; scalability and large-scale network analysis. Africa is fastest growing mobile phone market in the world. Incredibly smart kids in Africa hungry for this knowledge.

Note: this will be used by federal agencies to identify “terror” cells and predict criminal behavior.

Post a comment

You may use the following HTML:
<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>