> Once we have identified such lopsided data and understood its cause, how do you feed that back into your algorithm to correct for it?

This is tackled in the recently popular field of study called 'counterfactual inference'.

http://leon.bottou.org/talks/counterfactuals