Collaborative Ranking with Ranking-Based Neighborhood.

Recommendation system is a very important tool to help users to find what they are interested in on the web. In many commercial recommendation systems, only the top-K items are shown to users, and recommendation becomes a ranking task rather than a classical rating prediction task. In this paper, we propose a new collaborative ranking algorithm based on learning to rank framework in information retrieval. For a given user-item pair (u,i), we use Kendall Rank Coefficient as similarity metric to choose neighborhood for user u and use the ranking statistical information of item i from user u's neighborhood as the feature representation of pair (u,i). We apply LambdaRank to learn the ranking model and experimentally demonstrate the effectiveness of our method by comparing its performance with several collaborative ranking approaches. Moreover, we can address scenarios where users' feedbacks are non-numerical scores. © 2013 Springer-Verlag.