TORepair: Diffusion-Based Task-Oriented Error Repair Via Differentiable Bi-Level Optimization

Error repair in tabular data is crucial for downstream task utility. Existing methods either pursue data fidelity, potentially degrading task performance, or are task-oriented but require manual effort or extensive retraining, while lacking theoretical guarantees. In this paper, we propose TORepair, a novel framework that enables efficient task-oriented error repair through a differentiable bi-level optimization formulation. It features two co-evolving components, i.e., an inner optimization to update the task model with the latest repaired data, and an outer optimization to improve the repair model using the updated task model's guidance. Through a direct gradient-based feedback from the downstream task, the repair model is guided toward maximizing task performance. Meanwhile, to handle mixed data types and explore rich repair candidates, we adopt a conditional denoising diffusion mechanism that iteratively refines repairs in a continuous embedding space. The repair model is then jointly optimized with task-guidance for task utility and self-supervised learning for data fidelity regularization. We theoretically prove that our bi-level framework guarantees improvement over the no-repair condition. Extensive experiments on 11 real-world datasets demonstrate that TORepair enhances downstream task performance by 35% on average and outperforms state-of-theart error repair methods by 13% on average with efficiency, even surpassing clean data performance in several cases.