TOPOFE: Topology-Aware LLM-Guided Feature Engineering
A new framework called TOPOFE combines large language models with a topology-aware multi-island evolutionary algorithm to automate feature engineering for tabular data. Traditional AutoFE methods are limited by stateless generation and homogeneous search, converging to dominant patterns without discovering complementary feature compositions. TOPOFE introduces family-specialized exploration and adaptive pruning to overcome these limitations, enabling more diverse and effective feature transformations. The approach is formulated as a program synthesis problem, leveraging LLMs for generating feature programs beyond predefined operator libraries. The paper is published on arXiv with ID 2607.23286.
Key facts
- TOPOFE is a topology-aware multi-island evolutionary framework for LLM-guided feature engineering.
- It addresses limitations of stateless generation and homogeneous search in existing LLM-based AutoFE.
- The framework combines family-specialized exploration and adaptive pruning.
- AutoFE for tabular learning is formulated as a program synthesis problem.
- LLMs enable feature program generation beyond predefined operator libraries.
- The paper is available on arXiv with ID 2607.23286.
- The approach aims to discover predictive feature transformations from an exponentially large search space.
- TOPOFE prevents convergence to dominant transformation patterns.
Entities
Institutions
- arXiv