NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Bridging Expert Knowledge and Automated Feature Engineering via Self-Evolution
TL;DR
In high-stakes fields like clinical care and brand compliance, existing machine learning methods struggle to align with expert knowledge and qualitative criteria. FEST (Feature Engineering with Self-evolving Trees) was developed to generate interpretable features from unstructured data, combining semantic and deterministic approaches.
✦ Why It Matters
Engineers can leverage FEST to create interpretable machine learning models that align with expert knowledge in critical domains.
Key Takeaways
- Utilize FEST for generating interpretable features from unstructured data.
- Incorporate expert guidelines to refine qualitative criteria into operational features.
- Leverage the BrandGuide dataset for evaluating expert alignment in feature engineering.
- Aim for high semantic alignment in features to improve model accuracy.
Full Summary
High-stakes applications, such as brand compliance and clinical care, require machine learning models to be interpretable and aligned with expert knowledge. Traditional feature engineering methods often fail to meet these needs, particularly when dealing with unstructured data like text and images.
FEST (Feature Engineering with Self-evolving Trees) was created to address this gap by employing dual-stream feature generation, which includes both semantic and deterministic approaches, along with tree-guided iterative evolution. The results showed that FEST led to significant improvements, outperforming existing methods in 17 out of 20 classifier-task combinations, with an average accuracy gain of 4.2 percentage points.
Additionally, an evaluation using a large language model indicated that FEST achieved 60-80% coverage of expert-designed features, validated by human experts for relevance and clarity. By integrating expert guidelines, FEST effectively transformed qualitative criteria into operational features, enhancing accuracy by 6-12 percentage points across various brands.
This work introduces BrandGuide, a dataset that pairs expert-designed features with over 1 million assets, facilitating systematic evaluation of expert alignment in automated feature engineering.