I heard back from one of the authors of the unwanted traffic draft. In response I sent some comments it. (Progress! It's been a while since I've commented on a draft.) Some of my comments were based on the Wikipedia page on click fraud, which points to a paper on detecting coalitions of fraudsters – fraudsters who team up to generate fraudulent traffic on each others' sites. The research doesn't cover other types of click fraud, such as competitor-based or other-interested-party-based (e.g. investors who click on sponsored links just to drive up the revenues of the engines or ad networks they invest in). Still, it gives some bounds on how computationally difficult it is to detect this type of fraud, and some limitations on the methodology. IMO, this is valuable, even if it isn't perfect, because it gives advertisers an idea of how much they might have to pay (at a minimum) for a given amount of fraud protection. I think this type of work is more likely to stand up to technical criticism than vague claims about how much fraud is detected based solely on advertiser complaints.

The paper references research by Andrei Broder, who former AV folks will remember was Chief Scientist at one point. I wish I'd had an opportunity to work with him, and employ some of his techniques. At the very least, it would have given me more background and understanding of that type of analysis. Perhaps it would have made me a more attractive employee to someone, if not Yahoo! But I didn't get to work with him. :(