the program of the workshop is here in pdf. here are few notes:
orange folks presented a work in which they analysed how characteristics of individuals (e.g., age, gender) impact communication patterns. they found that 1) call duration and number of calls don’t change for males and females and don’t change across age groups; 2) the number of text messages drops dramatically with age. they are now trying to profile customers based on their local density in the communication network.
jp onnela and abersmann derived a communication network from four weeks of reciprocated calls and texts with the goal of studying the geographic patterns of social groups. to that end, they identified groups in the network using the modularity algorithm proposed by newman in 2006. as one expects, the geographic span (km) of a community increases with its size (number of nodes). interestingly, after normalizing geographic span with a simple null model and identifying geographic clusters using the k-mean algorithm, they are able to plot, for all clusters, their geographic span vs. their size. they observed that the geographic span increases linearly with size (for size below 30) and then flattens (no relationship between geographic span and size).
hossmann and his group at ethz presented a work in which they derive community structure from human contact networks.
gautier krings presented a cool analysis on temporal dynamics of social ties in a communication network (and corresponding temporal redundancy – great topic!).
belik of max-plank presented a work on the relationship between human movements and the spread of infectious diseases. they proposed a model of spatial epidemics that is more effective than the start-of-the-art model (i.e., the reaction-diffusion model). they found that global outbreaks are generated when the time spent by an individual in distant locations exceeds a threshold.
in a paper “a tale of two cities”, at&t folks analyzed mobility data in the cities of nyc and la and found differences (the daily ranges covered by travellers in the two cities are very different – but they haven’t normalized by population density though ;-)).
aharony of the mit human dynamics group is working on a mobile phone study called “friends and family”. they gave 100 phones to mit members who live in a residence hall with their families and administered them surveys at different times – monthly (20-30 minutes), weekly (5-10 minutes), and asynchronously (i.e., they call someone who has just terminated a phone call and ask him/her a couple of questions). the researchers studied the subnetworks of those people, which include the network of people who have kids, the network of those whose religion is A, the network of those living in floor B, network of those who have hobby C. the main idea is to study: 1) how influence (e.g., happiness) flows across those subnetworks; 2) how to nudge people (e.g., how to encourage people to loose weight) and how to measure the effectiveness of different intervention strategies (e.g., people that can be exposed to different views of their application store); 3) how friendship forms; 4) how people react if they are able to control their personal data.
finally, jean bolot of sprint gave a fantastic keynote titled “mining large-scale cell phone data”. the talk revolved around three main themes:
1. how mining cell phone data generates new business models
the telco operator could have a 2-sided business model: it could generate revenues from final customers (cell phone users) and from upstream customers (billboard managers, real estate companies). for example, telcos could rely on mobile phone data to:
1) dynamically price electronic billboards (whose pricing is currently fixed or is dynamically determined using a camera that counts the number of people passing by);
2) determine rent price for retail buildings (the rent of a store depends on its base of attraction – the distance people are willing to travel to visit the store)
the question is whether final customers could benefit (could get part of their money back) whenever their private information is used.
2. how mining cell phone data brings new life to old problem
social networks are boring, while mobile phone networks aren’t. for mobile networks:
. links reflect complex interactions (which could be represented by a time serie f(t))
. nodes move (nodes are function of time and position f(t,x))
. nodes come and go
. nodes pay to be in network (the link is f(price))
. not all nodes are in network (the network is incomplete)
. nodes belong to other social networks
by studying those networks, sprint folks found out that power laws aren’t ubiquitous as one would have thought. the distribution of the number of friends fits a double pareto log normal distribution instead. to model the addition of friends over time, they came up with a generative model of the form
Xt+1= Ft Xt
where Ft= Xt+1/Xt,
Xt+1= number of friends at time t+1,
Xt= number of friends at time t,
data shows Ft is lognormal.
(which is very different from the preferential attachment model)
3. cell phone data generates new applications
sprint researchers countered paging channel attacks by predicting which cell
tower a mobile phone user would be connected to [see mobicom paper in 2007]. it turns out that user mobility is highly predictable (H(x)= – sum pi log(pi), where xi is the number of calls made by the user through cell tower i, and M=sum x_i), especially if one separates weekdays and weekends. 14 days of data is enough to accurately predict user position (see NMI formula NMI(today, n days ago)= NMI(X,X-n) – to be found in their paper).