I committed MAHOUT-321, which changes how similarities are used as
weights in the calculations. Similarities are no longer shifted to the
range [0,2] from [-1,1]. This requires a little more care to
rationalize, in order to not produce predictions that are way out of
range.

It also copies logic into the user-based recommender from item-based
recommender, which will ignore any estimate based on only one
user-user similarity. The reasoning is that in a weighted average over
one data point, the weight is irrelevant, and that's not intuitive or
good.

It will change predictions. There are theoretical reasons it will
improve accuracy, and some empirical evidence too. However I think
this is a change worth continuing to play with and evaluate in
practice to ensure it's really helping.

Sean

Reply via email to