Research On Cold-start Problems In Collaborative Filtering Systems

The task of recommenders is to predict preferences on items for users or recommend items to users. Collaborative filtering is one of most popular technologies applied in recommender systems. This article is mainly to study and tackle cold-start problems in collaborative filtering methods. Cold-start problem is one of key challenges for recommender systems, especially in collaborative filtering recommender systems. It refers to recommenders cannot provide high quality recommendation for users (items) with few historical ratings. We consider three types of cold-start problems:cold-start system, cold-start users and cold-start items.We propose a feature-based regression model with baseline estimates noted as FRBE to solve coId-start problems. FRBE combines information filtering and collaborative filtering. We make full use of available user attributes and item features, and build a bilinear regression model. Plus, FRBE has baseline estimates of user and item effects.We also propose a method that combines trust-based recommendation and item-based collaborative filtering to tackle cold-start problems, and call it CoTrustWalker. It uses trust network in social network instead of neighbors in typical neighbor-based collaborative filtering. The method do not need any user attributes or item features, and can solve cold-start problems.After experiments, FRBE can provide high quality recommendation in cold-start cases. Compared to hot recommendation methods, the performance of FRBE are better. Because CoTrustWalker only relies on simple trust network, its performance is worse than that of FRBE.