Scalable recommender system that improves generalization

A major challenge for recommender systems is to generalize to cold-start prediction tasks, where no behavior data is available for the active user or the item. Content-based filtering is able to attack to this problem, while collaborative filtering ends up with accurate recommendations where high quality feedback is available. Considering the domain of a prediction task can vary, an ensemble learning-based hybrid recommender model is described. The combined model learns separate linear combinations from validation data sets representing each domain: high quality feedback available, user or item is unseen. The problem is illustrated by creating groups of validation and test data sets accordingly, and referring to three kinds of complementary recommenders: matrix factorization based, user demographics, and item content-based. Experiments demonstrate that using those separate validation data sets; the hybrid recommender model adjusts weights such that it converges to the individual recommender that performs the best on a domain.