2 posts tagged “netsci08”
Nicholas Christakis, Harvard, "eat drink and be merry, the spread of
health phenomena in social networks".
This talk is looking at the spread of desies throgh social interactions,
rather than other types of interactins. The main study was looking at
obesity using the Framingham Heart Study Social Network. This seems like
a very famouse social network health related study, so I'm not going to
go into detail about that, but the bottom line is that they were able to
construct the social interactions from this study by digging through the
huge paper archive. They were able to look at friend, relative and
co-worker ties.
The main study was looking at about 5k individuals out of 12k, taken
from 1973 onwards.
Nice, node ssize is related to a person's weight!
There is clear clustering of obese nodes in the network, now is this
clustering random or structured?
Well, it's more clustered than random.
There are a couple of reasons why this might be the case. It could be
that obese people like each other, people might be susceptible to local
factors, or there might be some kind of peer pressure.
By looking at time evolution the hope is that they might be able to find
'patient 0' for the obesity epedmic. OK, video is coming up now ...!
OK, looking at people getting fatter all over america from 1972 onwards,
I'm going for a run later!
The effect is not centered in one location, but it seems that it's an
epidemic that had multiple starting points in the network.
Looking at the directionality of ties of friendship helps you make
inferences about causes. Wow, if you are friends with someone who is
friends with you, and they get obese, you have 300% greater chance to
gain weight. Stay friends with thin people!
It looks like much of this is driven by social norms.
They also have gwo data from the network, that is really cool.
They can convert location to wealth, and can take this into account when
looking at the evolution of the network.
This data is really really cool.
No drop off in effect with distance, it is really the social tie that is
important.
They also looked at the effect of smoking, and were able to take this
into account.
So their working hyppothesis is that it might be the spread of behaviour
and habit, perticluarly shared behaviour, going runnning vs going for a
beer.
It might be the spread of an idea, the spread of what an acceptable body
size might be.
OK, that's pretty amazing, and you can tease a hugh amount of
information out of this study. Liklihood of quitting smoking, of how
that is effected by education, and friendship tie.
I have to say, there is not a lot of results that are amazingly
astonishing. They have food networks, like the bannana network and the
friend chicken network.
They are also looking at emotions. We know that emotions can spread
through groups, on diads. Could emotions spread hyper-diadically, and
over longer time frames?
There is strong clustering of happienss, and your happiness seems
coreelated with people who are outside of your direct social horizon.
Interestingly happiness does not spread in the workplace (I think that
was the point), but happy people have higher clustering and better
centrality in the network.
There seems to be a half life for catching happiness from your network,
this is about 6 months. There is also a strong local effect, you need
happy people to be within about two miles of you, and to be having happy
events happening to them every six months or so.
Ahh, you can look at smiling on facebook. Right, I gotta put up some
happy pictures on my profiles!
Ahh, thiness also spreads, but the reason they have been looking at
obesity is that this study is looking at the obesity epidemic. The
network shows you the magnification of the phenomena, not the cause or
origin of the phonomenon.
Interesting question, if you wanted to hire flight attendants who you
didn't want to gain weight, should you hire them based on the bmi of
their friends? Well, the answer is that in a workplace if a certain
behaviour begins to spread it is likely to have a network effect. The
flpiside is that you can use these network effects to more economic
effect by trying to promote certain behaviour through targeting core
groups in the workplace.
I'm in Norwich all this week attending Netsci08
http://www.ifr.ac.uk/netsci08/, the internatinal workshop and conference
on network science. It's a week long event, and broadly speaking it
looks like there are three types themes that are being discussed here:
biological networks, pure networks science and community detection in
networks, principlaly emergent networks of the kind we see in the
internet.
I'm twittering about the meeting using the tag #netsci08, but it seems
that I'm the only one out there in the twitterverse who is also at this
meeting. Not enough power in the lecture hall, and wifi is a little
ropey, but the conference is pretty good so far.
The talks on Monday were about some basics on network mathematics, and
on network science in the social sciences. I'll go back over my notes
and give a quick report on them when I get a chance to catchup, but the
discussion in the evening was pretty interesting, and the talk in the
morning touced on some very important topics.
The Tuesday morning tutorial is on economics and networks. The morning
model was very simple, and I think that's fair enough, but I got the
feeling that the level of the audience, at least on the side of the room
that I am sitting on, was high enough to have taken a bit more robust
model, so I got the feeling that there was some discomfot with the model
presented.
The after-coffee section is focussing on social influencers, now this is
interesting.
How is it that information flow is highly assymetric in the world?
The model is a mmulti-state model with differeing outcomes. Individuals
don't know the true state of the system, but they have beliefs about the
states. Sounds like a hidden markov model.
The model is stationary, and we want to see how the choices we make
change the beliefs that we have. Could be a bayseian network? Let's see.
What I am hoping to see from this model is how reccomendations can
travel throgh a network. There is a network of communication between the
network. The model can integrate dynamics, the dynamics of belief.
There is also feedback between actions and beliefs. The main result is
that as time goes by new information has less effect, and so beliefs
converge in the network. This is a consequence of Martingale's theorm.
The big question is whether we get optimal actions, and the big result
is that the ability to explore the action space and find the best action
is depenant upon the structure of the network. That is really
interesting.
Oh my God, someone has an OLPC machine in the audience, how cool is
that!!
Anyway, back to the talk. So this is indeed a Bayseian network. The
anti-intutive outcome from this model is that if you have to build a one
time only network that can't be changed later, then you have the best
chance of getting optimal behaviour if no one person has undue
influence, hoever I think that for online social networks there is a lot
of dunamics going on that can pull you out of local sub-optimal minima.