Evolving Demonstration Optimization for Chain-of-Thought Feature Transformation
New AI method uses reinforcement learning to evolve optimal prompts for data transformation tasks.
A team of researchers has introduced a novel framework called 'Evolving Demonstration Optimization for Chain-of-Thought Feature Transformation' to solve a core challenge in data-centric AI: feature engineering. Manually discovering effective mathematical transformations (like log, square, or binning) for raw data columns is notoriously difficult due to a vast combinatorial space. While Large Language Models (LLMs) have shown promise by using their internal knowledge to suggest transformations, current methods rely on static, pre-written examples (demonstrations) in their prompts. This leads to repetitive, low-diversity suggestions that aren't finely tuned for the specific dataset or prediction task at hand.
The new framework creates a closed-loop system that continuously evolves and optimizes these demonstrations. It starts by using reinforcement learning to explore and identify high-performing sequences of feature transformations. These successful 'trajectories' are stored in an experience library. For each new task, a diversity-aware selector picks the most relevant and varied examples from this library to construct the optimal prompt context for the LLM. This chain-of-thought style guidance steers the model toward generating transformation ideas that are proven to boost the final model's accuracy.
In experiments across diverse tabular datasets, the method consistently outperformed both classical automated feature engineering tools and other LLM-based baseline approaches. The system demonstrated greater stability compared to one-shot LLM generation and proved robust across different LLMs (both API-based like GPT-4 and open-source models) and different downstream machine learning evaluators. This represents a significant step toward making LLMs more reliable and effective partners in the crucial, but often tedious, data preparation pipeline.
- Uses reinforcement learning to build a library of successful feature transformation sequences (trajectories).
- Employs a diversity-aware selector to create optimal, evolving prompts for LLMs instead of static examples.
- Outperforms existing methods on tabular data benchmarks and works with both API-based and open-source LLMs.
Why It Matters
Automates and improves the critical data preparation phase for AI, leading to more accurate predictive models with less manual effort.