Some fun new terminology in the “let’s spam the recommender” business. Paul Lamere of Sun Labs reports on a talk by Bamshad Mobasher at Recommenders ’06 about attacks on recommendation systems.
[Dr. Bamshad Mobasher] outlined two basic types of attacks: shilling – trying to promote an item and nuking – trying to demote an item. These types of attacks are quite real. The social site Digg is under constant attack by shills trying to get their story promoted to the front page.
Bamshad pointed to an example when a loosely organized group who didn’t like evangelist Pat Robertson managed to trick the Amazon recommender into linking his book “Six Steps to a Spiritual Life” with a book on anal sex for men.
Bamshad suggest that one way to defend against shills and nukes is to create hybrid recommenders – recommenders that not only use social data but some inherent measure such as text or acoustic similarity. These types of systems are typically more robust than pure social recommenders.
This isn’t a new thing. Around 1998 at Infoseek someone was trying to sell us a system that used auto-categorization with discussion groups. As soon as they opened it up, some hostile users hijacked a Christian group to be in the Satanism category (or maybe Wicca, it was a long time ago).
At the start of this post I called recommender spam a business. I don’t know if anyone is making money on it yet, but it is one more opportunity for spammers. Your product gets a higher rating, they get paid. If you notice that sites suddenly require a login or even a subscription to post ratings, thank the spammers.