DECAF: Learning to be Fair in Multi-agent Resource Allocation