Eureka: An LLM-Driven Framework for Automated Feature Engineering in Enterprise AI

Summary

This paper presents Eureka, an LLM-driven framework for automated feature engineering in machine learning. It treats feature engineering as an agentic code generation problem where features are executable programs rather than static transformations. The framework has three stages: (1) an Expert Agent fine-tuned via SFT produces structured feature design plans, (2) an LLM Feature Factory translates plans into Python code using chain-of-thought reasoning, and (3) a Self-Evolving Alignment Engine uses Reinforcement Learning (GRPO) with dual-channel rewards to improve code quality. Evaluated on 7 public benchmarks across healthcare, finance, and social domains, Eureka outperforms traditional AutoFE and LLM-based baselines. In a real-world deployment at Alibaba Cloud for GPU resource demand prediction, Eureka improved demand fulfillment rate by 16% and reduced computing resource migration rates by 33%.