344 — Detecting Influence Campaigns in Social Networks Using the Ising Model

Bots are a rampant issue on Twitter, and it doesn’t really seem that Twitter is spending a lot of energy internally dissuading these botnets or their creators.

Not that it’d be hard for them to do something; it’s not too difficult to detect these bot accounts.

This paper is an example of one technology that does so very efficiently, classifying whole cohorts of accounts as bot or not-bot at a time. It does this by solving a minimum-cut problem — the task of minimizing the error when performing a linear bisection of an embedded population — embedded in the first place using the Ising model of magnetic kinetics.

It is conventional for bots to mount influence campaigns that use many bot accounts to target users all simultaneously. It’s not effective, obviously, for bots to target other bots (especially those in the same botnet) and so bots tend to interact more with users than with other bots. But this makes it easy to identify them: Bots exhibiting this “heterophily” — preferring human interactions over bot interactions — makes them conspicuous. In fact, this “magnet”-like interaction means that they can be represented as a magnetic interaction, and solved using a multidimensional Ising model.

The results of this are subjected to a minimum cut, the results of which maximally segment bots and humans in large batches (in contrast with other approaches which must address each account individually).