A Novel Recommendation Relevancy Measure for Collaborative Filtering.

Recommendation systems help people find their potential interests. In recommendation algorithms relevancy measures play an important role. Current relevancy measures often employ only user-item rating data or combine with contextual information to obtain related users or items. However, in some specific situations, these measures may not guarantee high accuracy or sufficient candidates. This paper solves these problems by proposing a novel recommendation relevancy measure, which indicates how worthy an item can be recommended to a user. In this paper, each interaction between a user and the recommendation system is regarded as a behavior represented with several features. The relevancy measure is achieved with a series of stepwise calculations and combinations on these features and behaviors. We evaluated the effectiveness of this measure against three other popular measures with a public dataset extracted from a commercial search engine. The experiment result shows that it can generate more recommendable items and achieves both high recall and precision.