What kind of normalization do you have in mind? I can imagine many variations on what it does now, each of which has some merit.
Without ratings, the classic simple neighborhood algorithm is a bit inapplicable: it relies on a weighted average of ratings, where similarities are weights -- but there are no ratings. So you have to rank based on some function of the weights only (similarities). I think it merely sums now, which is the simplest thing. On Wed, Jul 13, 2011 at 4:23 PM, Steven Bourke <[email protected]> wrote: > Hi - > > It looks like the boolean recommenders would benefit from some form of > normalisation in the predictive stage of the recommendation algorithm. An > item that is prominent in a users neighbourhood may lose out simply because > of a high similarity between two single users. > > I'm assuming there is a reason this hasn't been added previously, anyone > know why? >
