Dual-Label Learning With Irregularly Present Labels