Bridging Expert Knowledge and Automated Feature Engineering via Self-Evolution

Researchers have introduced FEST, a method that translates qualitative expert guidelines into auditable machine-learning features, targeting high-stakes domains where model decisions must be inspected and aligned with human oversight [1]. The system, formally called Feature Engineering with Self-evolving Trees, combines dual-stream feature generation — one semantic, one deterministic — with semantic deduplication and tree-guided iterative evolution to extract interpretable features from raw text and images [1]. In evaluations spanning brand classification, content authenticity detection, and stress detection, FEST led in 17 of 20 classifier-task combinations, posting a mean gain of 4.2 percentage points over the strongest baseline across five classifiers [1]. A central challenge in regulated settings is that existing automated feature-engineering tools target tabular inputs and cannot operationalize qualitative criteria such as “maintain professional tone” into precise, measurable features [1]. FEST addresses this gap by refining expert-supplied guidelines into operational features, which improved accuracy by 6 to 12 percentage points on average across brands [1]. An LLM-as-judge evaluation found that FEST achieved 60 to 80 percent coverage of expert-designed brand features at strict semantic-alignment thresholds, a result corroborated by a human expert study that rated the features highly on relevance, clarity, and actionability [1]. To enable systematic benchmarking of expert alignment in automated feature engineering, the authors released BrandGuide, a dataset pairing expert-designed features with more than 1 million assets across 2,683 brands [1]. High-quality labeled datasets of this kind are typically difficult and expensive to produce because of the time required to label data, making BrandGuide a notable resource for supervised and semi-supervised learning research [6]. The work appears on arXiv, the open-access repository that hosts e-prints across physics, computer science, and related fields and that, as of late 2024, receives roughly 24,000 submissions per month [10]. The paper’s abstract page also surfaces community-built tools through arXivLabs, a framework launched in 2020 that allows collaborators to develop and share experimental features — such as citation explorers and recommender systems — directly on the site [9]. Those integrations are designed to uphold arXiv’s values of openness, community, excellence, and user data privacy, with third-party collaborators granted only minimal, anonymized data for functionality [9].