Towards A Scalable Solution for Improving Multi-Group Fairness in Compositional Classification