Yes I think I understand what you're getting at and the examples. Loss function here is just the 'penalty' for predicting a rating near to those of dissimilar users and far from those of similar users?
If I read correctly, you think that a 'weighted average' (with negative weights in numerator and denominator -- yeah I don't like using the absolute value in the denominator, conceptually) plus capping is an intellectually sound way of handling this situation. And I think I am convinced by this way of rationalizing it, and so am no longer scared of negative weights. Let me then create a patch for this. On Tue, Feb 23, 2010 at 7:21 PM, Ted Dunning <[email protected]> wrote: > Weights can't be negative and still be weights. You can have large > (positive) weights on negative training examples (aka "not like this"), but > you can't really have a negative weight. >
