Personalized paper recommendation in online social scholar system

This paper presents a practical paper recommender system, which aims to provide personalized research paper recommendations to users within an online social scholar system. As an online recommender system, there are three basic problems we need to tackle: 1) How to formalize and solve the recommendation problem; 2) How to achieve real time recommendation; and 3) How to interact with users. In our work, we take the personalized paper recommendation as a ranking problem with respect to users´ research interests, and employ a supervised learning to rank approach to solve the problem. However, most previous learning to rank methods rely on manually labeled training data which are both expensive and limited in size. We propose automatical training data construction by mining the existing large scale academic network, and extract various heterogeneous features for learning. With the learned model, we conduct real time personalized recommendation based on our novel efficient candidate generation approach. In addition, update recommendation is employed to interact with users according to different types of user feedbacks. Finally, we demonstrate the effectiveness of our system by both offline and online evaluation.