Table 3.

In predicting trolling in a discussion, features relating to the discussion’s context are most informative, followed by user-specific and mood features. This suggests that while some users are inherently more likely to troll, the context of a discussion plays a greater role in whether trolling actually occurs. The number of binary features is in parentheses.

Feature Set	AUC
Mood
Seasonality (31)	0.53
Recent User History (4)	0.60
Discussion Context
Previous Posts (15)	0.74
Article Topic (13)	0.58
User-specific
Overall User History (2)	0.66
User ID (45895)	0.66
Combined
Previous Posts + Recent User History (19)	0.77
All Features	0.78