Recommending based on rating frequencies

Since the development of the comparably simple neighborhood-based methods in the 1990s, a plethora of techniques has been developed to improve various aspects of collaborative filtering recommender systems like predictive accuracy, scalability to large problem instances or the capability to deal with sparse data sets. Many of the recent algorithms rely on sophisticated methods which are based, for instance, on matrix factorization techniques or advanced probabilistic models and/or require a computationally intensive model-building phase. In this work, we evaluate the accuracy of a new and extremely simple prediction method (RF-Rec) that uses the user's and the item's most frequent rating value to make a rating prediction. The evaluation on three standard test data sets shows that the accuracy of the algorithm is on a par with the standard collaborative filtering algorithms on dense data sets and outperforms them on sparse rating databases. At the same time, the algorithm's implementation is trivial, has a high prediction coverage, requires no complex offline pre-processing or model-building phase and can generate predictions in constant time.