Peer evaluation through cross-efficiency based on reference sets