I've been working on extracting new features for rated items by analysing users' rating behaviour. These features are incorporated into classifiers and evaluated in terms of precision/recall. I'm interested in comparing my features with other similar research efforts . Are there any state-of-the-art recommendation methods where rating information is analysed in order to produce new features?
Take a look in Taste Project. Avaliable at: http://taste.sourceforge.net/
Thanks a lot for your answer.
I've noticed that this project is a part of Mahout Machine Learning project which contains a number of implemented recommenders.
https://cwiki.apache.org/confluence/display/MAHOUT/Recommender+Documentation
I'm not aware of any recommender in that project that extracts new features for items. All implemented algorithms estimate ratings for items.
Is there something that I'm missing?
excuse me!
I dont understand your question! how can you extract features from rating?
can you describe more :)
e.g in movie lens data set we just have movieID, UserID and Rate(1-5)
you can join movieID with movies.dat file and extract its categories.
by using matrix factorization methods you can extract movie features and use cosine similarity measure to find similar movies.
I suppose matrix factorization is an approach of extracting features for items from user-item ratings matrix. Is the following paper a representative work for such techniques?
http://dl.acm.org/citation.cfm?id=1608614
Btw, I also make use of ratings in order to calculate user interest levels in items attributes such as genres. From these levels I produce new features for rated items.
Hi,
matrix factorization is fine but a linear approach. If you are interested in nonlinear models take a closer look to localized variants of relevance learning vector quantization (GRLVQ). GRLVQ classifies data vectors according to predefined classes while weighting the attributes according to their contribute for classification. Extensions of GRLVQ takes into account correlation information between the attributes (matrix learning -> GMLVQ). Frequently, both methods are comparable to SVMs!
The koren et al. paper is a good resource for Matrix factorization approaches.
We have argued in previous papers that the features used in MF could be assigned psychological meaning and we have used them to diversify recommendation sets with success. See an early version of this paper in these workshop proceedings:
http://ceur-ws.org/Vol-811/paper3.pdf
- Badges
- Science topic