From Hackathon to production: building a large scale recommendation engine (exchange.causes.com)

Most of what I learned was from Programming Collective Intelligence (http://amzn.com/0596529325). If you're in the Bay Area, Noisebridge has a self-taught machine learning class that I went to for a while (great group, I just couldn't make the time commitment).

Hm, I haven't. This suggests that Mahout might not work well for datasets on the scale we're dealing with (working toward 1 billion membership rows):

http://lucene.472066.n3.nabble.com/Mahout-not-giving-recomme...

But that's also an older article, and things might be better. Mahout might be worth checking out.