Evolutionary Feature Engineering for Structured Data

Large language models are increasingly used as open-ended search operators in evolutionary optimization. We introduce Evolutionary Feature Engineering (EFE), a framework that uses LLM-based evolution to discover preprocessing programs for structured data. EFE represents transformations as Python programs with a standardized fit/transform interface and refines them using dataset context, summary statistics, and validation feedback. For time-series forecasting, EFE-Time learns invertible dataset-specific normalizations that improve time-series foundation models, reducing MASE, WQL, and MAE by at least 3\% on average and up to 19\% on COVID-Deaths. For tabular prediction, EFE-Tab evolves compact, interpretable feature programs that improve or match prior LLM-based feature-engineering methods, with especially strong gains for decision trees. Overall, EFE shows that LLM-based evolution can improve structured-data model performances.