Flexible collaborative filtering: a bayesian approach

Abstract

Recommendation systems seek to predict the "rating" or "preference" that a user would give to an item. We propose a new method for collaborative filtering that allows flexible recommendations to users through Markov chain Monte Carlo algorithms. With this approach, one can draw from the predictive posterior distribution and use it to produce point estimators. After convergence, each sampled value can be used as rate prediction. Our approach allows fast results to be produced since it does not require waiting for the simulation of a full chain before making predictions. This is not only welcome by the users themselves but also helps with the learning mechanism of the algorithm. Also, one of the biggest concerns of this study was to create an algorithm that is scalable. To do so, we suggest optimization steps to bypass matrix inversion. That way, even under high dimensional scenarios, the computational time of our algorithm is satisfactory. Finally, an application to the Movie Lens data set Harper and Konstan (2016) is presented as na illustration.