The project aims at using different recommendation methods for different kinds of real world data like rating matrices, images and text, using Deep Learning and Optimization.
The project is distributed into 8 chapters:-
- Given a sparse rating matrix, the goal is to come up with an ordering of items for the users. Regularized Matrix Factorization techniques are usually effective because they allow us to discover the latent features underlying the interactions between users and items.
- Order preservation above Matrix Factorization has been shown superior to Matrix Factorization where we tried to fit the ratings and also preserve the relative ordering of ratings by transforming the rating matrix using learned monotonic functions σ’s. A practical drawback of these algorithms is getting good estimates of hyper-parameters involved in these algorithms, a naive approach being grid searching.
- An alternative way to estimate hyper-parameters is by using Emperical Bayes methods and using them in the optimization algorithm.
- A Content based recommendation systems that make suggestions based on similarities between user and item feature vectors taken from the user and item factor matrices.
- We from here on use different Unsupervised Deep Learning methods to learn important high level features of item representations.
- Matrix factorization can be further used to express user and movie features for user-movie-similarity based arguments for recomemndation.
- Documents or Natural Languages are known to be a discrete combinatorial system. Comparison in continuous space becomes less relevant because of its discrete nature, therefore learning good high level continuous features becomes important before comparison and recommendation.
- Images are very high dimensional representations of objects which have inherent localized structures that requires feature learning before relevant comparision.
We use unsupervised Deep Learning to learn such high level features in Documents, Images and User-movie ratings. These features become useful when we further use vector space models to get good recommendations.
The Thesis report can be found here