REFINE: Enhanced Reinforced Feature Generation via Semantic-Guided Exploration

The quality of machine learning models strongly depends on the quality of the input data, making feature engineering a crucial step in the machine learning pipeline. Traditional feature engineering often requires extensive domain expertise and manual effort, which can be time-consuming and challenging to scale. Furthermore, with the proliferation of ML-powered systems, especially in critical contexts, interpretability and explainability have become essential to ensure trust and transparency. To address these challenges, we introduce KRAFT, an automated feature engineering framework that employs knowledge graphs with reinforcement learning to enhance both predictive performance and interpretability. KRAFT employs a hybrid AI approach that combines a neural generator for feature transformation with a knowledge-based reasoner leveraging description logics to assess feature interpretability. The generator is trained using deep reinforcement learning, optimizing both predictive accuracy and interpretability simultaneously. By systematically exploring and validating feature transformations within a structured knowledge framework, KRAFT ensures that the generated features are not only effective for prediction but also align with human-understandable concepts. Extensive experiments on real-world datasets demonstrate that our approach significantly improves model performance while maintaining a high degree of interpretability. This work highlights the potential of combining symbolic reasoning with data-driven methods to advance interpretable ML, offering a promising direction for trustworthy AI systems.