Bridging Expert Knowledge and Automated Feature Engineering via Self-Evolution

Researchers introduce FEST, a machine learning system that automatically engineers interpretable features from unstructured text and images while aligning with expert knowledge. The method outperforms existing approaches across brand compliance, content moderation, and clinical tasks, and the team releases BrandGuide, a new dataset of 1M+ assets with expert-designed features for systematic evaluation.

FEST addresses a critical gap in interpretable machine learning for high-stakes domains where model decisions must be auditable and grounded in expert judgment. Traditional feature engineering methods struggle with unstructured data and cannot operationalize qualitative expert criteria into quantifiable features. This work combines semantic analysis with deterministic feature generation and tree-guided evolution to bridge that divide, achieving 4.2 percentage point improvements over baselines across multiple classification tasks.

The research emerges from growing regulatory and organizational pressure for AI transparency in sensitive sectors. Brand compliance, content moderation, and clinical care increasingly require systems that practitioners can inspect and validate. As enterprises face scrutiny from regulators and stakeholders, the ability to produce features that align with documented expert guidelines becomes competitive advantage. FEST's approach—generating both semantic and deterministic features, deduplicating them, then iteratively refining through tree structures—creates a reproducible path from raw data to auditable decision logic.

The release of BrandGuide represents a significant infrastructure contribution, providing the first large-scale dataset pairing expert-designed features with real-world assets across thousands of brands. This enables the ML community to systematically measure expert alignment in automated feature engineering, previously an underspecified problem. For organizations deploying ML in regulated industries, FEST reduces the friction between model performance and interpretability requirements. The 6-12 percentage point accuracy improvements when seeded with expert guidelines demonstrate that incorporating domain knowledge enhances both accuracy and transparency.

• →FEST achieves expert-aligned feature engineering with 60-80% semantic coverage of human-designed features across strict evaluation thresholds.
• →Dual-stream generation (semantic and deterministic) combined with tree-guided evolution outperforms baselines in 85% of tested classifier-task combinations.
• →BrandGuide dataset enables systematic evaluation of expert alignment in feature engineering, addressing a previously unmeasured capability.
• →Integration of expert guidelines into FEST improves accuracy by 6-12 percentage points, demonstrating synergy between automation and domain knowledge.
• →The approach scales to multimodal inputs (text and images) while maintaining interpretability requirements for high-stakes deployment scenarios.