SIGMA: New LLM Framework for Metadata-Free Automated Feature Engineering
arXiv:2608.17948 introduces SIGMA, a scalable constant-context optimization framework for metadata-free automated feature engineering (AutoFE) built on large language models. The paper identifies two limitations in current LLM-based AutoFE: reliance on unavailable semantic metadata, and trajectory accumulation risking context-window overflow or unstable generation that leads to local optima and feature duplication. SIGMA replaces semantic information with SHAP values to deliver task-aware signals for group feature generation. It also adopts the EXposed-feature Implicit Trajectory (EXIT) approach. These elements aim to improve scalability and stability across long-horizon optimization.
Key facts
- The paper is titled 'SIGMA: SHAP-Guided Implicit-Trajectory Generation for Metadata-Free LLM-Based AutoFE'.
- It is an arXiv preprint with ID 2608.17948v1 (cross announcement).
- SIGMA is a scalable constant-context optimization framework.
- LLM-based AutoFE faces challenges with semantic metadata availability.
- Trajectory accumulation can exceed context windows.
- Without trajectory, generation may become unstable, causing local optima and duplicate features.
- SIGMA uses SHAP values for task-aware signals instead of semantic information.
- SIGMA adopts the EXposed-feature Implicit Trajectory (EXIT) approach.
Entities
—