Retrieval-Augmented Feature Generation for Domain-Specific Classification

Domain-specific classification often operates with limited data and a small set of available features, while manually creating additional informative features requires substantial expertise. Retrieval-Augmented Feature Generation (RAFG) identifies associations among existing features, retrieves relevant external knowledge, and uses large language model reasoning to generate and validate interpretable candidate features. Experiments on medical, economic, and geographic datasets show that the generated features are meaningful and improve downstream classification performance compared with existing feature-generation approaches.

📄 In Proceedings of the 25th IEEE International Conference on Data Mining (ICDM 2025), pp. 943–952.
🔗 Paper Link (IEEE)