Tony Jebara from Columbia University and Sense Networks, “Mobile Phones Reveal the Behavior of Places and People” on Wednesday morning at ETech 09.
[Missed a few minutes.] Stuff about making social networks in Facebook etc. How do we capture these networks in the real world? Sense networks using data from iPhones, TomTom, cellphones, etc. Lots of data being collected all the time.
They monitored people who commuted into the San Francisco financial district. A correlation [looked very small] between how early people arrived at work and the state of the Dow Jones. Monitored the hours people go out in the evening. Again a correlation with the Dow Jones [couldn’t understand the graph at a glance so not sure how much there was].
CitySense.com — real-time density of users at every street corner. Poisson models find the most active bars and restaurants. Gives you a “sixth sense” of where people are in the city. So you can see if lots of people are gathering somewhere. [Oh, only in San Francisco.]
What’s next? Filter it for where people like you are. Different colours for different types of people, a heat map of “where’s my crowd?” They currently have 20 different tribes. Colour bars at the top of the screen indicate which tribe you belong to most at the moment, changing over time depending on your behaviour.
They build a network of people from location data. Like Facebook’s network of people but without the self-reporting. You’re connected to someone because you’re co-located with them. And building links with people who hang out in similar kinds of places, eg, two people who visit Starbucks in two different cities at similar times of day have a connection.
The only data they have are GPS tracks for people. They need to translate the raw latitude and longitude into semantic data; what does it mean to be in a particular location at a particular time. They build a network of places, identifying similar places no matter where they are. To identify this similarity they use three kinds of information: the commercial and demographic activity (using government databases), plus flow analysis.
Flow analysis data: If several people arrive in two places from similar sources and then leave to go to similar destinations, this indicates similarity between the two places. So you can tell that two bars in different parts of the city are probably the same kind of place as they attract people who go from and to similar locations.
Commerce data: Standard Industrial Categorization (SIC) Code. Assigns a number to a building depending on its commercial use.
To work out the profile of an individual they analyse the Flow, SIC and demographic data for the places they go. He shows a table with rows for hours (168 for every hour of the week), and columns of different flow, SIC and demographic types. The table cells show the percentage chance of a person being in a particular column in a particular hour of the week. At this point in the process the GPS data is thrown away. The matrix is adjusted over time as the user changes their behaviour. They only store the current matrix for a person.
People can be connected by their similar behaviours. Uses:
- Churn — eg, a mobile phone company sees one person has left their network, they might then promote themselves to connected people to stop them leaving too.
- Advertising —
- Marketing —
- Collaborative Filtering —
- Demogrpahics —
If people go to car dealerships often enough they can be identified as potential purchasers. These people can be broken down into high or low-end car purchasers depending on which dealerships they go. [This is all starting to sound a bit evil, rather than fun.]
They get very good predictions of how someone will respond to ads. [I think that was right.]
[For all that stuff about how anonymised the data is earlier on, it sounds like if they have a client that’s say a phone company, they are able to tell the phone company about an individual’s behaviour. Which, of course, the phone company need to to provide a useful service to their clients, but it does sound rather evil (or “commercial”, depending on your point of view).]
One example of a less “depressing” application (in response to a question): if you go to a new city they could recommend places you would like, rather than just recommending exactly the same places to everyone.
They claim around four million users.