My class, Network Epistemology, went virtual. I'm still planning to do a twitter thread about each week.
This week we looked at two similar, but distinct, phenomena: homophily and conformist bias. We focused on their epistemic effects, how do they effect learning?
This week we looked at two similar, but distinct, phenomena: homophily and conformist bias. We focused on their epistemic effects, how do they effect learning?
Homophily is that similar people tend to be more connected in a social network. Think: people who have similar political views tend to hang out together.
Conformist bias is that people tend to change their opinions (or stated opinions) to match the opinions of those around them.
Conformist bias is that people tend to change their opinions (or stated opinions) to match the opinions of those around them.
The first paper we looked at was by @ben_golub and Jackson. This paper looks at how homophilous networks perform on a particular epistemic task. Are networks where people are more or less segregated into different groups better or worse at that task?
academic.oup.com/qje/article-ab…
academic.oup.com/qje/article-ab…
They use the repeated averaging or "pooling" model of group learning, and show that while in the long run homophilous groups do the same as non-homophilous groups, in the short term homophilous groups are much worse. Not only do they take longer to converge to the best answer...
... but they can also initially head in the wrong direction. That is, they don't uniformly approach the best answer to the question, but might take a detour before eventually getting there.
The paper is exceptionally interesting and does a lot more than my short summary covers. (They also defend a novel measure of homophily.)
We looked \at repeated averaging models in a previous week, and many of the same questions and concerns arise there.
We looked \at repeated averaging models in a previous week, and many of the same questions and concerns arise there.
The second paper we read on homophily is very different, it looks at the degree to which existing networks (in this case twitter) are homophilous. This really interesting paper is by Sullivan et al (including @moral_psych).
philarchive.org/archive/SULCRS
philarchive.org/archive/SULCRS
They look at a month of tweets on vaccine effectiveness, and analyze the network formed by people retweeting one another. They find that the network is incredibly segregated by opinion, very little information goes back and forth between the two opposing viewpoints.
The conclusion is not terribly surprising to those of us on twitter, although I was amazed at how *extremely* homophilous it was.
The concern they raise, quite reasonably, is that this limits how "wise" we might thing crowds are... at least on this issue and on this venue.
The concern they raise, quite reasonably, is that this limits how "wise" we might thing crowds are... at least on this issue and on this venue.
In class, we had some concerns about the timescale of the analysis. I was curious, for instance, if other epistemic groups -- like scientists -- might look similarly homophilous if only investigated for a short time.
In addition, there was worry about how people were included in the sample. Perhaps only those who are already in like-minded groups are likely to tweet regularly about vaccines. Maybe the non-homophilous part of the graph gets left out of the sample.
But despite those concerns, we thought the approach was interesting, and I personally would love to see more work on topic-specific social network structure -- I think this is a great avenue for empirical study.
Lastly, we looked at a paper by @__mnml (Mohseni) and Williams on conformity instead of homophily. They begin with a model from Smith and Sørensen, where individuals get private signals about the state of the world and make public declarations.
aydinmohseni.com/wp-content/upl…
aydinmohseni.com/wp-content/upl…
But in addition to wanting to be correct, they also assign some negative value to disagreeing with their friends. So, I might think that the Steelers will be terrible next year, but be unwilling to say that since my friends disagree.
In their model, the agents will *eventually* converge to all saying the truth. But in the short run network structure matters a lot. Whether more or less connected networks are good depends on some underlying parameters of the model -- which is itself pretty interesting.
But the one universal is that the "star" network, where one individual is connected to everyone else, is very bad. This stands in contrast to other models (we'll look at next week) where the star network is the *best* network. 
In the model agents are Bayesian learners, and this had one assumption that we questioned. An agent knows what their neighbor's neighbors are saying. If I see you say "Steelers are going to win" I know what your friends (who aren't my friend) are saying about the Steelers.
But the models is an exceptionally interesting method for incorporating conformity into traditional "rational" economics models, and the analysis is lucid and very easy to follow. I recommend you check out the paper.
Overall, this week was interesting. It represents a collection of analyses -- using very different techniques -- of a decidedly irrational (or arrational) behaviors: homophily and conformity. In all cases, such behaviors were clearly counterproductive.
But there are various ways to mitigate their harm, either by ensuring that we give groups time to overcome the harm or by ensuring that the groups are not overly centralized (like the star graph).
