Scientist: Influencer Theory Is Bogus


But marketers aren't interested in that?

What I started to realize after several of these conversations was the whole influencer theory is actually more of a rhetorical device than a theory. It's not like people actually have an explicit theory of who is influential and they go out there and they use that theory to decide who to target. It's more that they do whatever they do. They throw parties or they give away free samples or they advertise in particular publications that brand themselves as reaching an influential audience and run some campaign in their normal manner. And if it works, they say, "We reached the influencers." And if it doesn't work, they say "We didn't reach the influencers" or  "We weren't able to make them help us." In that sense, it's just a rhetorical device to help you explain the randomness that you actually experience in the world. I thought people actually believed this and in believing it they actually used it and that's bad because people should be thinking more systematically. So I was sort of trying to say to them "OK, if you want to find the influencers, I'll tell you how to do that.

And if it turns out that you can't find them and in fact you're better off seeding things randomly, for example, then that's great because you stop wasting your time and energy. But the reaction to that was I think informative that in fact they weren't doing that to begin with...The irony I think is that targeting ordinary people is probably what marketers are doing and if something takes off they say "Oh, we reached the influencers." So it's one of these impossible-to-falsify theories because who you identify as an influencer is always after the fact.
What are you working on at Yahoo?

We are still interested in social influence. What we're trying to do now is use network data in a non-experimental manner to try to see -- if you look at a diffusion of a particular product over a social network just using observational data-you can ask a question about if this product gets taken up by a group of people and subsequently their friends adopt it -- it's really like a real-life simulation of the models we were running. In a large network, you're dropping a seed and asking how much do each of these seeds grow until they stop. Now we have a bunch of trees, diffusion trees, and some of them are very small and a few are very big. And what you'd like to be able to do is say "Can I predict the size of the diffusion trees based on what I know of the seed?" This is the influencer's hypothesis, that there's something special about the seed that predicts the size of the tree.

We ran that in simulation results and now what we're trying to do is replicate that using empirical data, which is not quite the same as dropping it in the real world, but it's the best we can do at this point. We are still pursuing this question of what predicts success. You can throw into that model both the attributes of the people and the thing itself, the thing that's diffusing and you can throw in attributes of the network. That's where it might get interesting. If you found for example that you can't really predict what product is going to take off and you can't predict which individual is going to be responsible for spreading it, but there are features of the network which are conducive or not conducive to spreading. That would really be interesting. Then you'd have a true network theory of diffusion. If you can go out and measure certain features of the social network you can actually optimize your targeting around those features. From that perspective, the influencer hypothesis is just the simplest network statistic that you can measure, like who had the most links or who had the most Twitter followers. Let's call them the influencers. So a strategy of picking the most highly connected nodes is like the very simplest network measure that you can think of. You might say "Well, that doesn't really work because people with high degrees of [connections] are followed by people who aren't that interested in what they have to say." Or the people who follow them are people who no one else is interested in reaching. So you can reach a bunch of people, but it never goes anywhere.
You might find a mix works better?

The insight that we got out of the simulation models was the network attributes that was conducive to diffusion is: Easily influenced people influencing other easily influenced people. If you get a long sequence of those individuals, you can propagate for many steps. Now the interesting question is does anything like that rule come out of the real-world data?