LLM-TDA: A LLM-based Tabular Data Augmentation Framework

Tabular data augmentation is a critical technique for improving machine learning performance in data-limited scenarios. However, existing methods face inherent limitations in balancing quality and efficiency. Filter-based approaches tend to introduce low-quality or irrelevant features due to the lack of task-specific feedback, while wrapper-based methods incur prohibitive time overheads necessitated by massive downstream model training. Fundamentally, both paradigms rely solely on statistical properties, neglecting the rich semantic information in tabular data. To address these limitations, this paper proposes LLM-TDA, a novel framework that leverages the semantic reasoning capabilities of Large Language Models (LLMs) to optimize the feature augmentation process. LLM-TDA employs a coarse-to-fine pipeline: first, a table-level analysis rapidly prunes semantically irrelevant tables based on business logic; second, a fine-grained feature selection mechanism fuses semantic understanding with statistical assessment to identify high-value features. Extensive experiments on real-world datasets demonstrate that LLM-TDA significantly outperforms state-of-the-art baselines in downstream task performance. Notably, it achieves this accuracy while delivering a 10-fold to 14-fold improvement in computational efficiency compared to wrapper-based methods, providing a robust and scalable solution for automated feature engineering.