My own best candidate for using side information, of which context is just one source, is the latent factor log-linear approach described in Menon and Elkan's paper. I am part-way into an implementation of this, but it will not be integrated into the recommendation framework at first. As soon as the freeze lifts, I wil commit partial results.
Here is the paper in question: http://arxiv.org/abs/1006.2156 2010/10/12 Matthias Böhmer <[email protected]> > Hi there, > > I have been to ACM RecSys and presented a paper at the workshop on > context-aware recommender systems about appazaar, a context-aware > recommender system that suggests apps for the Android phone (please > drop me an email if you are interested in that topic, you can find the > app on the market). > > Using the IDRescorer basically follows a post-filtering approach -- > i.e. applying context-information after e.g. a collaborative filter. > The method of adding extra dimensions for context mentioned by Steven > --- who by the way also gave an interesting talk in the CARS workshop > ;-) --- follows a model-based approach, where you integrate > context-information into the model itself (see the chapter on > Context-aware Recommender Systems on the forthcoming Recommender > Systems Handbook for a discussion on contextual pre-filtering, > contextual post-filtering, and contextual modelling approaches). There > was an interesting paper presented by Karatzoglou et al. on > "Multiverse recommendation: n-dimensional tensor factorization for > context-aware collaborative filtering". I would really like to see > this algorithms implemented in Mahout. Context-awareness is a really > important topic for recommender systems. > > Best, > Matthias > > 2010/10/12 Steven Bourke <[email protected]>: > > Hi - There was an entire track dedicated to context recommender systems > at > > this years ACM Recommender Systems, including a challenge using the movie > > pilot dataset > > > > There is a substantial amount of research out there about using > contextual > > features in a recommender system. A pretty mainstream approach is to add > an > > extra dimension to your matrix when calculating recommender scores. If > you > > pop onto scholar.google.com and check out context recommenders or tensor > > matrix you'll find some useful literature. Just adding simple rescore > values > > would probably make these additional attributes fall into the domain of > > context based recommenders. > > > > On Tue, Oct 12, 2010 at 12:04 PM, Sean Owen <[email protected]> wrote: > > > >> Yeah this is embodied in the "IDRescorer" class which lets you > >> influence the final recommendations however you want, for just this > >> sort of reason. > >> > >> On Tue, Oct 12, 2010 at 11:55 AM, Sebastian Schelter > >> <[email protected]> wrote: > >> > Hi everyone, > >> > > >> > I have some non-release-related offtopic questions ;) > >> > > >> > I've attended an interesting talk last week given by the CTO of > >> > moviepilot.de (a German movie recommendation platform). The talk > >> included a > >> > concept which he called "context-aware" recommendations, which means > that > >> > you not only recommend items that a user might like, but you also > factor > >> in > >> > the users current context. The example in the talk was that a romantic > >> > evening needs other movie recommendations than an evening with a > couple > >> of > >> > guys drinking beer. > >> > > >> > I found this concept very appealing and I thought about whether and > how > >> this > >> > could be accomplished with our current recommender framework. > >> > > >> > My idea would be to define some content-related rules like "a movie > >> tagged > >> > with the category action is not suited for the context romantic > evening" > >> and > >> > use these rules to create some kind of context-aware Rescorer that > only > >> > selects items matching the rules from all recommended items. > >> > > >> > Would this be a viable approach? > >> > > >> > --sebastian > >> > > >> > > >
