|
|
|
|
|
|
by rohanphadte
3155 days ago
|
|
|
Hey, I'm a creator of the tool written in the article. The model looks at hundreds of features of each profile, not just a few features such as patriotic, inflammatory language, and retweets as you suggested. As a result, the model has predicted profiles with very sketchy behavior, such as accounts of normal people that were compromised and become political propaganda accounts a a few years later. We wanted to bring up these accounts and the analysis behind it so we could show the techniques behind these organizations who run these bot-like accounts. We offer our own analysis of these accounts here: https://medium.com/@robhat/an-analysis-of-propaganda-bots-on... |
|
|
The Wired take definitely left me cynical; it's all too easy to write a human-interest piece about ML approaches that don't actually work. But this is a much more concrete explanation of results, and I'm intrigued.
If you don't mind, could you offer any more clarification on your test/training set? I see the Medium piece talks about wanting to avoid selection bias from hand-classification, but the Wired summary just described hand-selecting 100 "ground truth" bots and then adding their followers. How did you know the followers were also bots? And how did you try to ensure the bots you selected were a reasonable sample?