[ak217]

Sounds like a challenge :)

First, this doesn't have to be for each pageview. Crudely, you want every story to have a vector in some giant hyperspace characterizing it, and every user to have a vector precomputed offline based on their previously shown affinities, and take dot products.

This is not necessarily easy, or computationally very cheap, but the payoff can be pretty big.

I believe LinkedIn runs something similar to this every night.

[rorrr]

I'm not sure what you mean exactly, but hey, if you can pull it off, it might be successful.

Just think of the number of comments up/down voted every millisecond, and that's just one the dimensions. The scale of that matrix would be enormous.