Abstract:
Collaborative filtering (CF)-based recommender systems predict what items a user will like or find useful based on the recommendations (active or implicit) of other members of a networked community. In spite of more than ten years of research, there is little consensus on state-of-the-art knowledge regarding CF predictive algorithms. There are many barriers to synthesis of the significant quantity of available published research on CF algorithms. We present results from an empirical study that attempts synthesis on popular CF algorithms and use this study to illustrate some key challenges to synthesis in CF algorithm research. In response to these challenges we propose the development of publicly maintained reference implementations of proposed CF algorithms and empirical evaluation procedures and we
introduce CoFE, a public software framework with the goal of jumpstarting the building of these reference implementations. Finally, we demonstrate how CoFE was used to implement a high-performance nearest-neighbor-based algorithm that scales to arbitrary numbers of users.
