(Hmm, I don't know why it doesn't post to the mailing list. We get a message about moderating everything. I'll copy it to the list now.)
In #1, you describe the usual user-item preference matrix. Yes it's sparse. I guess you could make up pseudo-items like genre in the matrix, yes, if you had explicit ratings for genres, and then perhaps over-weight these in the computation. How do you use this matrix? there is no algorithm that doesn't use this in some way! that's exactly what a DataModel provides. You can use demographic information in a few ways. Yes you could use it as part of a filtering process (e.g. a dating site would never recommend males to heterosexual males) or a boosting process (maybe you somehow know the user likes shoes, so you make shoe recommendations extra high-priority). These processes are not really 'learning' or collaborative filtering but can be necessary to implement business logic. You could also use user-user similarity based on attributes as the basis of a UserSimilarity metric and use it in a user-similarity-based recommender. Of course it requires you've already figured out some method of determining similarity based on attributes, and that's up to you to make up. Sean On Thu, Apr 5, 2012 at 10:16 PM, <[email protected]> wrote: > Hi Sean, > > I'm a new user to Mahout and am currently using the 1m ratings MovieLens data > file. I've seen your responses on several forum threads and was hoping you > could answer a couple of questions I had. I'm aware that as PMC chair you > probably don't have much time to answer. If that's the case, could you please > point me in the direction of someone who might be able to help? I've posted > online to the forum directly, but my message is underscored by the "This post > has NOT been accepted by the mailing list yet.: > > #1. ItemBased using CF and content-based > I've been thinking of a way to incorporate content-based recommendations > (using additional attributes of an item, in this case a movie) by utilizing a > utility matrix where the rows correspond to users and columns correspond to > movies. The individual entries (say entry(i,j)) are the preferences (ratings) > expressed by user i for movie j. I know you can include additional columns, > say genre, and identify a value for that entry based on the movies rated by a > particular within that genre type. > > This matrix would be pretty sparse. But I'm not sure *how* to combine this > with the out-of-box recommendations provided in Mahout. How will I do this if > my data model (a matrix) is different from DataModel? > > #2. Adding in demographic information > I'd like to recommend movies to users, and not users to users, so I'm not > sure how I can incorporate user "attributes" (e.g. their age group, > occupation, etc) without doing this. The Libimseti dating recommendation > makes sense - there, we're recommending users to users. In my case, does it > make sense to just filter based on a particular e.g. age group, and make a > recommendation based on this filtered group? (I guess this rests on the > assumption that my filter will return users that are more similar to each > other because they are in the same demographic category (in this case age > group).) Or is there a better method of doing this? > > Any advice would be very helpful. Thanks in advance. > > Anita Mehrotra
