Reducing offline evaluation bias of collaborative filtering algorithms