Bob Carpenter: If you automatically interpret a link from page A to page B as a sign of trust, and you don't have any other data, then you can model trust and lack of trust, but not distrust. One solution that's becoming more popular is to analyze larger link structure communities and figure out which ones are positive links and which ones are negative using not only local information about a page but aggregate information about the page's place in the community. For instance, political blogs often link to articles they don't like. They often provide linguistic clues in the surrounding text and anchor text. The techniques for analyzing this are the same ones Nicolas Nicolov was telling you about for your book. In aggregate, if a blog mostly links to right-wing blogs, links to left-wing blogs have a higher likelihood of being negative. So you can use these bits of information to mutually constrain your inferences. (P.S. You should really be asking them about [partial] singular value decomposition, the workhorse of the Netflix prize, and the generalization of eigenvalues/vectors to arbitrarily shaped matrices!) Dec. 7, 2009, 4:21pm