Improving K-nearest-neighborhood based Collaborative Filtering via Similarity Support

Collaborative Filtering (CF) is the most popular choice when implementing personalized recommender systems. A classical approach to CF is based on K-nearest-neighborhood (KNN) model, where the precondition for making recommendations is the KNN construction for involved entities. However, when building KNN sets, there exits the dilemma to decide the value of K --a small value will lead to poor recommendation quality, whereas a large one will result in high computational complexity. In this work we firstly empirically validate that the suitable value of K in KNN based CF is affected by the number of the totally involved entities, and then focus on optimizing the KNN building process for providing high recommendation performance as well as maintaining acceptable KNN size. To achieve this objective, we propose a novel KNN metric named Similarity Support (SS). By taking SS into consideration, we designed a series of strategies for optimizing the KNN based CF. The empirical studies on public large, real datasets showed that due to the improvement on KNN construction brought by SS, recommender optimized by our strategies turned out to be superior to original KNN based CF in terms of both recommendation performance and computational complexity.