Another month has gone by which means another paper to cover!
This month, instead of covering my usual stomping grounds of biology or chemistry, I decided to look into something a little bit more related to my work in venture capital: social networks!
The power behind the social network concept goes beyond just the number of users. Facebook’s 500 million users is pretty damn compelling, but what brings it home is that by focusing on relationships between people rather than the people themselves, social networks turn into a very interesting channel for information consumption and influence.
This month’s paper (from Damon Centola at MIT Sloan) covered influence – specifically, how different social network structures (or “topologies” if you want to be snooty and academic about it) might have different influences on the people in the networks. More specifically, it asked the question of what social network would you expect to be better able to influence behavior: one which is more “viral”, in the sense that connections aren’t clustered (i.e., I’m as likely to be friends with my friend’s friends as people my friends don’t know), or one which is more “clustered” (i.e., my friends are likely to be friends with one another).
It’s an interesting question, and I found this paper notable for two reasons. First, its the most rigorous social networking experiment I’ve ever seen. Granted, this isn’t saying very much. Most social network/graph studies are observational, but I was impressed by the methodology and the attempt to strip out as much bias and extraneous factors as possible:
- The behavior being tested was whether or not they would sign up and re-visit a particular health forum. This forum had to be valuable enough to get people to use it (and actually contribute to it), but also unknown and inaccessible to the rest of the world (as to avoid additional social cues from the user’s “real world” social network).
- The author (and I do mean one single author: pretty rare these days for a Science paper as far as I know) created different social graphs which were superficially identical (same number of users, same number of contacts per user) but had the different network structures he wanted to test(one structure had subgroups of tightly inter-connected users, the other structure had random connections scattered across the network). The figure below shows one example of the network structures: the black lines show connections between people. On the left-hand-side is the highly clustered social graph – the individual users are only connected to people “next to them”. The right-hand-side is the more “viral” social graph, where users can be connected to any user across the social network.
- The users made profiles (with user name, avatar, and stated health interests), but to preserve anonymity (and limit the impact of a person’s “real world” social network on a user), the user names were blinded and users were not allowed to directly communicate (except in an anonymized fashion through the health forum) or add/remove contacts
- However, whenever a user’s contacts participated in the health forum, the user would be notified.
The result was a somewhat bizarre and artificial “network” – but its certainly a very creative (and probably as good as it can get) means of turning social networking studies into a rigorous study with real controlled experiments.
Second, the conclusion is interesting and has many implications for people who want to use social networks to influence people. Virality may be a remarkably fast way to get people to hear about something, but the paper concludes that virality does not necessarily translate into people acting. The author conducted 6 different trials with slightly different network topologies (number of users ranged from 98 to 144, number of contacts per user ranged from 6 to 8). The results are in the graph below which shows the fraction of the users who joined the forum over time. As you can see, the clustered networks (solid circles) had much higher and faster adoption than the “viral” networks (open triangles):
Why would this be? The author’s standing theory is that while “viral” networks might be faster at disseminating information (e.g., a funny video), clustered networks work better at driving behavior because you get more reinforcement from your friends. In a clustered network, if you have one friend join the forum, chances are the two of you will have a mutual friend who will also join. At a very basic level, this means you get the same cue to join the forum from two of your friends. In a un-clustered network, however, if you have one friend join the forum, the two of you are less likely to have a mutual friend, and so you are less likely to receive that second cue.
Does this matter? According to the study, someone who had two contacts join the forum was ~75% more likely to join than someone who only had one contact join. And, someone who had four contacts join was ~150% more likely to join than someone who only had one contact. While this effect rapidly diminshes with more contacts (having five or six contacts join made relatively little difference compared with four), its a powerful illustration of quality vs. speed in a social network – something which is also borne out by the fact that while only 15% of people who only had one contact join returned to the forum, 35-45% of users who had multiple contacts join did.
This was definitely a very impressive and well-designed study. While it would be fair to attack the study for its artificiality, I don’t really think there’s any other way to systematically strip out the biases that are intrinsic to most observational (not a controlled experiment) studies of social networks.
Where I do think this was lacking (and maybe the researcher has already teed this up) is the black-and-white nature of the study. What I mean by this is while I find the argument that network clustering helps drive greater behavior plausible, I think there needs to be a more rigorous/mathematical conception – how “clustered” does a network need to be? If a network is overly clustered, then it loses the virality which helps to spread ideas more quickly and widely – is there an optimal balance somewhere in the middle? Also, the paper only dug, on a very superficial level, into how network size and the number of contacts per user might impact this. I think further experimental and mathematical modeling/computational studies would be nice to really flesh this out.
Paper: Centola, Damon. “The Spread of Behavior in an Online Social Network Experiment.” Science 329 (Sep 2010) – doi:10.1126/science.1185231
(Image credit – social network diagram) (Figures 1 and 2 from paper)