Collaborative Filtering with Constraint Aspect Model

Nicolas Delannay,
phD student,
Machine Learning Group, UCL
supervisor: Prof. M. Verleysen.

Abstract
Collaborative filtering is the method of automatically finding patterns in
the interactions between a set of users and a set of items. In particular,
collaborative filtering is one of the approaches used to recommend new
products in e-commerce. The idea is that users who tend to have similar
behavior in the past should also have similar behavior in the future.
One of the approach for collaborative filtering is known as the Aspect
Model. This model tries to represent the behavior of a user by a mixture of
typical profiles. The model is handle by probabilistic framework and the
profiles are estimated by an EM algorithm. The Aspect model is also called
probabilistic latent semantic in the field of text analysis.
While seductive in its principle, the Aspect Model is prone to overfitting
due to the large number of parameters to estimate. Also, the model fails to
deliver profiles easy to interpret because they are defined over the
complete set of items. Interpretation can be simplified by adding a
variable representing the activity of an item with respect to a profile. To
circumvent overfitting, we propose to couple these activation parameters
with the rating distribution parameters. We show that the corresponding EM
algorithm looses its analytical updates but can be handle efficiently with
iterative gradient ascent steps. The benefit of the constraint model is
presented on the movielens database.