Follow up from last message I forwarded:
---------- Forwarded message ---------- From: Sean Owen <[email protected]> Date: Mon, Apr 13, 2009 at 6:02 PM Subject: Re: problems with taste To: 郑楠 <[email protected]> OK, I think this will work as you want, with one change. UserNeighborhood needs to use modelPaper, not model. In general I think you will get weird results by mixing DataModels - the code generally assumes all components use the same one. But in this particular instance I think it may work. I note that 'model' contains little data so you may have trouble establishing any correlations between users. Users need to overlap in at least two items to compute a correlation. Also you are using a neighborhood size of 1, which limits recommendations to the items that that one user rates, that the current user does not. This is why you are not getting k=4 recommendations. I would suggest you use Pearson instead of Spearman to start. I would also suggest you avoid a preference inferrer (though it makes the above problem worse - need more data in 'model'). Finally I would forget the CachingRecommender for now. Lastly, in modelPaper, if all prefs are 1 you can simply omit the ',1' in the file (I am assuming you use FileDataModel somewhere to read the file; it will handle this). It will be a little more efficient but it hardly matters at small scale. At large scale it may. On Apr 13, 2009 12:52 PM, "郑楠" <[email protected]> wrote: To be honest, my dataset is some bookmarks from CiteULike. I want to use the user-tag Matrix to compute the similarity between users, and use this result to recommend papers to these users (using user-paper Matrix). Can I realize this using taste? Thank you!
