"We could also take other data into account, like the user who submitted the article, and generate features indicating things like the karma of the user..."

I'm wondering how helpful could that be in prediction though? Would it actually help if I wish to predict how many upvotes my headline would get, and I add my karma as a feature? I think in fact such features would degrade generalisation performance, as they stand in like placeholders (when training), i.e. high karma users are correlated to higher probability for a "hit story".