Mitigating the objective mismatch for Recommendation Systems: Boosting self-supervised representation learning with instance-based return optimization