Assigning reliability values to recommendations using matrix factorization

Abstract

Providing a reliability value to each prediction and recommendation is very important in current recommender systems: Users should know which recommendations are reliable and which ones are risky. Despite its growing importance, research into collaborative filtering reliability has rarely been developed in the model-based area. This paper explains a matrix factorization-based architecture and method that provides a reliability value to each prediction/recommendation. The reliability values obtained have been put to the test, and, when applied, they show improvements in prediction and recommendation quality in different recommender systems; additionally, they provide a range of values that are understandable to users.