Hacker News Reader

Yep this is 100% true (i run shepherd.com).

Within our data books like lord of the rings and harry potter are nearly impossible to map for "books like" because they are connected to so many other things. I am working now to fine tune our model, but it has been an interesting challenge.

Yeah basically what I think you're saying is that in a weighted graph, if there are edge weights which are orders of magnitude larger then the average it throws off certain models. Like nearest neighbor.

Basically just prune the top and bottom %1 of weighted edges to get an appropriate average. Would be my guess for a fix.

IIRC you could do pretty okay in Netflix's recommendation contest by ignoring what an individual likes and just recommending the stuff that "everybody" likes.