Yes, in one of your data sets, I noticed that all the preference values were "1". This indicates to me that you really don't have a notion of the strength of the preference between users and items. There is an association, or there is none. I call this, somewhat wrongly, a "boolean" preference.
In this case, you can use faster and lighter versions of the components you are currently using, which are specialized for this situation. To try this, first use a copy of your data file which omits the final ",1" on every line. You don't need it. Instead of using PearsonCorrelationSimilarity, try BooleanLogLikelihoodSimilarity. Remove the PreferenceInferrer (these don't work so well anyway in my experience) Then use BooleanUserGenericUserBasedRecommender as your recommender implementation. For such a small data set, it is already extremely fast. But if you had a great deal more data, you would see a big difference. You may even find this approach, which ignores preference data, gives better results. On Thu, Jul 16, 2009 at 11:16 AM, Laya Patwa<[email protected]> wrote: > Thank you so much guys for discussing the problem of mine. I am getting the > recommendations now! > You mentioned something about improving the performance in one of the mails.
