A Task-Driven Interpretable Data Cleaning Framework Based on LLM

Addressing challenges in data-intensive applications including low data quality, limited adaptability, and poor interpretability of traditional cleaning methods, this paper proposes TR-Agent, a task-driven, large language model (LLM)-based framework for interpretable data cleaning. Designed for multi-source heterogeneous tables with incomplete metadata, TR-Agent integrates four core components: tabular knowledge extraction, error detection, error repair and report generation. Structured cleaning reports enhance transparency and traceability, while feedback from downstream tasks dynamically refines cleaning strategies to ensure both data cleanliness and practical utility. Experiments on multiple real-world datasets show TR-Agent consistently outperforms traditional methods and baselines. Notably, it achieves an 8% higher F1 score on the Titanic dataset and a 7% improvement on the Breast Cancer dataset compared to HoloClean. These results validate TR-Agent’s ability to generate high-quality, utility-driven data for real-world cleaning scenarios.