Understanding Bad Fake Follower Scores via @OwenJones84
By Rob | 16 Feb 2015, 12:00pm | Category: Social Media
On Thursday and Friday a couple of articles were published in the press about our App. One about David Cameron and one about @OwenJones84 and they highlighted some interesting points about Fake Followers on Twitter.
The articles were overtly political in nature, so I don't wish to delve into their specifics or take any sides. Also it should be noted, that on a personal level, I share neither man's politics so I regard myself as a neutral. What I would like to discuss though are the nature of bad Fake Follower scores, how our app works and give a clearer insight into the data.
According to our OwenJones84 Faker Page -- updated Feb 14th -- Owen has a faker score of 84%. This is of course very high and suggests that something unusual has occurred. It does not mean that Owen has purchased any Fake Followers. To put this in context Owen should have no need to. He's a very popular Guardian journalist, political firebrand and has a highly engaged Twitter account, so it's not surprising he has a lot of followers. In addition the amount of spam on Twitter should not be underestimated, it's a huge problem that Twitter cannot solve without ruining their network.
So why might our app register such a bad score for Owen? Essentially it comes down to the way the Twitter API works. To analyse 1,000 followers requires us to make at least 6 requests to Twitter. If you combine this with the way that Twitter limit requests per 15 minutes there is a natural block to the amount of data we can analyse in a short period of time. Our aim is to return scores within 30-120 seconds so currently we sample up to 1,000 records over the first 35,000-50,000 records we access. It should also be noted that Twitter return follower data most recent follow first.
What does this mean in Owen's case? My theory would be that someone has purchased Fake Followers for Owen's account. It is incredibly easy to do with sites like this for any Twitter handle. The account has then been analysed and because the most recent 50,000 or so followers are fake our app has generated a terrible score. Please note this is just my theory based on previous experience, I'm not pointing the finger of blame anywhere. Nor am I suggesting the first part and the latter part of my theory are linked in any way.
Now of course this does open up our app to the accusation that our algorithm can be gamed. It can, but only in extreme circumstances like Owen's. According to our data over 95% of Twitter users have less than 50,000 followers and 65% have less than 2,000. So for the great majority of users our scores will be perfectly accurate and reflect their true Faker Status.
In addition to this we also have a Deep Dive tool that over the course of a couple of days analyses far more data than our live tool. This is so we can validate the scores of larger accounts for those interested. On Friday I set this off on Owen's account to give us a better understanding of his data. The results are as followers...
- We sampled 16,154 records accross his follower base of 286,848 followers
- It returned scores of 26% Fake, 29% Inactive and 45% Good
This we believe is a much more accurate score and backs up our theory that Owen has a relative normal following for someone with his social status.
We hope this information clears things up and gives everyone a better understanding of how our app works.