Our Method


We use a novel metric that does not measure the direct links in a social network, but instead measures the amount of common friends (or followers) between two accounts. This co-following property is interesting, because it measures similarities between two items of popular culture, without the need for a direct link or endorsment between the tow. For example, a lot of Bernie Sanders supporters might be interested in Tesla, however Sanders himself has not expressed any particular inclination towards a particular brand of cars.

In the above diagram, we give an example on how the method works. Here, the score of a lifestyle account (e.g. a popular brand of soda) is derived from the common followers between the account and the two policitians Sanders and Portman.


Our measurement uses the political scores of congresspeople as a ground truth. Here’s how it works. Our procedure begins with 553 current and recent members of the U.S. Congress whose political alignment is derived from their roll call voting using a widely accepted metric developed by political experts Poole and Rosenthal. These scores range from -1 (very liberal) to +1 (very conservative). We then identify their 5 million followers and assign each follower a political score that is the weighted average of the scores of the Congress members that each one follows. Finally, we identity all the Twitter accounts that are also followed by these 5 million followers of Congress. Any account that is followed by several followers of Congress is then assigned a political score based on the weighted average of the political scores of their followers who also happen to follow members of Congress.

For example, many people who follow Bernie Sanders and Elizabeth Warren also happen to follow Prius, while followers of Ted Cruz and Marco Rubio follow Harley Davidson. Thus, Prius ends up being scored at the "blue" end of the spectrum, while Harleys are scored at the "red" end. Scores of lifestyle accounts are then assigned, in essence, by looking at the fraction of their followers overlapping with politicans' accounts.

We can use this method to score any of the accounts that you follow, as long as these accounts are also followed by enough followers of Congress that we can obtain a reliable estimate of the relative interest among the followers of liberals and conservatives in Congress. You can think of these scores as measuring whether a randomly chosen follower is more likely to be liberal or conservative. You can then use our analysis to see if your lifestyle interests (as indicated by who you follow) tend to overlap more with the interests of liberals or conservatives. For example, if your interests in movies, music, and sports tend to overlap more with liberals, this might indicate that you are more likely to also be liberal, or vice versa for conservatives.

Related Work

A similar method has been applied to data extracted from Reddit by Hessel et al.