A Novel Collaborative Filtering Using Kernel Methods for Recommender Systems

Recommender systems form an essential part of e-business systems. Collaborative filtering (CF), a widely used technique by recommender systems, performs poorly for cold start users and is vulnerable to shilling attacks. Therefore, a novel CF using kernel methods for prediction is proposed. The method is called Iterative kernelbased CF (IKCF), for it is an iterative process. First, mode or mean is used to smooth the unknown ratings; second, discrete or continuous kernel estimators are used to generate predicted ratings iteratively and to export the predicted ratings in the end. The experimental results on three real-world datasets show that, with IKCF as a booster, the prediction accuracy of recommenders can be significantly improved especially for sparse datasets. IKCF can also achieve high prediction accuracy with a small number of iteration.