I've just re-read section 4.2 exploring the user-based recommender - and the role of the similarity measure is their, front and centre!
cheers lee c On 26 October 2011 12:39, Sean Owen <[email protected]> wrote: > A-ha. I should elaborate then. The essence of the item-based algorithm > is estimating prefs as weighted averages of other prefs. The weights > are similarities. This depends on having prefs to average in the first > place in the data model. But it doesn't depend on whether the > similarity value uses ratings or not. Those weights are just weights, > wherever they come from. > > The recommenders that operate without ratings don't actually compute a > weighted average anymore -- it doesn't make sense. They compute > something else to rank on, but it's no longer an estimate pref > actually. That's why AAD doesn't have real meaning there. > > But in either case you're welcome to use log-likelihood similarity for > example which does not depend on pref values at all. It's just > supplying a value which is used as a weight in the first instance, and > something else in the second instance. > > On Wed, Oct 26, 2011 at 12:35 PM, lee carroll > <[email protected]> wrote: >>> AAD is not valid for comparison when you're not using >>>ratings in your *recommender*. It's nothing to do with your similarity >>>metric. >> >> The penny drops. >
