think of more heuristic troll-detection features

[corcra]

So far I have:

• % tweets insulting
• % tweets from conversations
• average conversation length

Need more suggestions for features. They don't necessarily need to be very informative; if we get training data a classifier can sort that out.