Abstract
The goal of recommender systems (RS) is to provide personalized recommendations of products or services to users facing the problem of information overload on the Web. The most popular approaches to retrieve the most relevant items for a user are collaborative filtering (CF) recommendation algorithms and in particular in recent years a number of sophisticated algorithms based, e.g., on matrix factorization or machine learning, have been proposed to improve the predictive accuracy of RS.
In our recent work, we proposed a novel recommendation scheme called RF-REC, which generates predictions simply by counting and combining the frequencies of the different rating values in the usual user-item rating matrix. The scheme has some key advantages when compared with more sophisticated techniques. It is trivial to implement, can generate predictions in constant time and has a high prediction coverage.
In this paper we propose extensions to our method in order to further increase the predictive accuracy by introducing schemes to weight and parameterize the components of the predictor. An evaluation on three standard test data sets reveals that the accuracy of our new schemes is higher than traditional CF algorithms in particular on sparse data sets and on a par with a recent matrix factorization algorithm. At the same time, the key advantages of the basic scheme such as computational efficiency, scalability, simplicity and the support for incremental updates are still maintained.