Political Experiments and Information Cascades
Source: https://civic.mit.edu/blog/erhardt/learning-from-political-experiments-and-information-cascades-on-facebook
In early 2014, Lada Adamic, a data scientist at Facebook, conducted research on what metrics are most useful in the prediction of information cascades. To begin her research, she examined how political news propagated throughout Facebook. She grouped news into four buckets based on their source:
Potential – the combined total of all the news being shared by all of your friends
Exposed – the news that actually showed up in your news feed
Selected – the news that you actually clicked on
Endorsed – the news that you liked
As it turns out, people generally tend to have a balance of liberal and conservative friends, regardless of their political leaning, however Facebook has more liberal leaning people than conservatives, which means the potential bucket of news generally has a slight skew towards liberal news. The exposed bucket is then largely determined by the endorsed buckets of your friends, which is the beginning of the information cascade. If one of your friends like something, it is more likely to be exposed to you, which makes you more likely to select it (Lada discovered that news in the selected bucket was proportional to the news in the exposed bucket, that is, everything in the exposed bucket is equally likely to be clicked on regardless of content), and once you select something you are more likely to endorse it, which continues that information cascade by then causing that source of news to show up in your friends’ news feeds. This is directly related to what we learned in class about information cascades, in that once one or two people endorse something, it becomes dramatically more likely that other people will end up endorsing it.
The concept of information cascade that we talked about in class also explains an unusual phenomenon that Lada came across in her research. She found that, although Facebook has more liberals than conservatives, the distribution of liberal and conservative news was almost equal. She also saw that conservatives were less likely to like liberal news than liberals were likely to like conservative news. This meant that the size of the endorsed buckets were higher for conservatives than it was for liberals. This means that, even though there are more liberal news sources in the potential bucket, since conservative news has a higher endorsement rate, it evens out and causes there to be an equal distribution of liberal and conservative news.
Lada then goes on to talk about how timing, network structure, content, origins, and veracity all play a role (in addition to endorsement rates) in determining if something will trigger an information cascade, and she cites research done in part by Professor Kleinberg (“Can Cascades be Predicted?” http://arxiv.org/abs/1403.4608). Ultimately she says that all the factors are important, however the speed at which the news initially propagates is the most important, even more important than if the information is true or not. Once again, this coincides with what we learned in class, namely how information cascades can easily cause incorrect information to spread.