[PaulHoule]

It is a tough problem no doubt, largely because of the unbalanced sample size.

People who work on anomaly detection in finance (anti-fraud) usually look at it as a "characterize a normal transaction and reject anything that is far from the center" problem as opposed to a "classify transactions as good or bad" problem.

Would someone succeed or fail? I think it could go either way. Are you talking about the general problem or the problem for some particular environment? (ex. Netflix certainly does not need to solve the general problem)