Mitigating cold-start problems in drug-target affinity prediction with interaction knowledge transferring