Social Network Construction and Analysis Using Sensor Networks

Michael Farry , Samuel Madden & Jonathan Pfautz (Charles River Analytics, Inc.)

We are developing methods for the construction and analysis of social networks using small, radio-equipped, battery-powered sensor devices known as "motes," which obtain data about the environment around them. Our goal is to develop probabilistic models that, given traces of sound, temperature, and vibration data from sensors on people, can accurately predict whether those people were interacting with each other. We plan to build these probabilistic models from sensor data traces where interactions are known and use those learned models to detect the presence of unknown interactions in other traces.  Predicting these sorts of interactions with mote devices is challenging because data rates on motes are quite low – at best, a few samples per second. Such low-rates are of interest because they support very discreet devices and very low-power modes of operations that will allow data collection for weeks or months.

Interpreting readings from sensors in a meaningful way is a challenge. We plan to use techniques from the machine learning community – in particular, Bayesian Belief Networks (or BBNs) – to help interpret these readings. BBNs can be used to model the probability that two people are interacting as a conditional probability distribution that depends on the values of sensors on each person. We use probabilistic inference over BBNs to allow predictions about interactions to be made despite missing, uncertain, or noisy sensor data. To train the system (e.g., learn the probability distributions) we have captured video and sensor traces from several scripted scenarios with actors. Using the video data, we can label the traces with information about when interactions were occurring; we can then use this tagged data to train our models and test their predictive power.

The other major aspect of this research involves learning social networks. We plan to develop a programmatic system that can construct a social network given inferences from models built over sensor data. We seek to answer important questions about sensor-tagged individuals. For example,  given two parties, what is the frequency and duration of their communication? Have they been in contact recently? What means do they use to communicate? What is the best way to categorize their communication? Is their relationship an informal friendship, or is it a more businesslike relation? Is there any hierarchy established by their communication? It's unclear at this point what granularity we may be able to achieve in interaction classification, and part of this research is identifying that what granularities are feasible from low-data rate networks like motes.

We believe the ability to automatically construct social networks from low-data rate sensors could be useful in a number of arenas, ranging from espionage to better understanding the flow of information or the nature of relationships in social groups.