Identifying Twitter Bots

About a year ago, I was doing some data mining on Twitter to do sentiment and network analysis on tech related subjects. Along the way, I found some interesting patterns and some insight into the bot world.

Here in this network graph, the size of the bubble indicates the number of edges (ie. a retweet or tweet towards that account) connecting them to another account. This is taking a look at the search results of hashtags related to coding. Take a look at this picture, notice anything strange?

With this particular account, you see that the majority of the accounts interacting with it have the name “Bot” in it: ie. _FunBot, PythonExpertbot, TheDeveloperBot. These are all bots created to retweet when certain tags are mentioned. Notice that interestingly NONE of these bot accounts interact with anyone besides the account they are retweeting. A high number of bots in tech related subjects make sense, as users in those fields are more likely to have knowledge on how to use the twitter api to create such bots.

Bots or not?

The above bots have no qualms with being found. They explicitly mention that they are bots. What would an ideal bot behave like if they wanted to stay hidden?

Here is a network consisting of these tweeting at uber_support. The way this network was constructed was getting all tweets sent to that account and also colouring the communities that exist using k-clique identification. K-Cliques are interesting because they can help us identify communities/threads by identifying users that are “adjacent” to each other. In this case every user is adjacent to atleast 3 users in that clique. Compared to the bots, this network has users interacting with users besides the main account. However, in this case, it’s very difficult to find out which is a bot due to the fact that the users this network interact with each other like normal people do.

In the picture below we identify the cliques and the clique “leaders” those that are tweeting the most to that account in their respective cliques. If bots acted in this way it would be very difficult to identify them.