Scholarlink

<p>Transparent feature pipelines in federated learning must balance interpretability, privacy, and system cost. The synthesizes methods for feature engineering and feature selection in centralized, horizontal federated, vertical federated, and hybrid settings. The evidence base includes 74 studies published during 2020–2025 in conferences, journals, and preprints. Retrieval, screening, and extraction follow a PRISMA-style workflow and a structured codebook that records data partition, feature operations, privacy mechanism, threat model, interpretability channel, and reported metrics. A thematic synthesis of the 74 selected studies is organized into four strands: interpretable automated feature engineering, federated feature pipelines, privacy and fairness controls, and systems and tooling for reproducibility. Evidence indicates that automated feature search paired with explicit parsimony constraints and pipeline-aware explanation mapping can increase transparency while limiting compute. Reporting of expression complexity, input-space fidelity, and stability under resampling remains inconsistent. In vertical federation, compact representation exchange and selective re-evaluation can reduce rounds and bandwidth, but pipeline-composition leakage and verification cost are rarely measured end-to-end. Differential privacy supports feature selection and explanations, but pipeline-level budget accounting and validation of synthetic-data feature importance remain open challenges. Provenance-rich pipelines and meta-feature profiling support reproducible deployments, yet standardized scorecards and head-to-head benchmarks remain scarce.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Scalable, interpretable, and privacy-preserving feature engineering for federated learning

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Vijay Kumar Kambhampati,
  • K. B. V. Brahma Rao

摘要

Transparent feature pipelines in federated learning must balance interpretability, privacy, and system cost. The synthesizes methods for feature engineering and feature selection in centralized, horizontal federated, vertical federated, and hybrid settings. The evidence base includes 74 studies published during 2020–2025 in conferences, journals, and preprints. Retrieval, screening, and extraction follow a PRISMA-style workflow and a structured codebook that records data partition, feature operations, privacy mechanism, threat model, interpretability channel, and reported metrics. A thematic synthesis of the 74 selected studies is organized into four strands: interpretable automated feature engineering, federated feature pipelines, privacy and fairness controls, and systems and tooling for reproducibility. Evidence indicates that automated feature search paired with explicit parsimony constraints and pipeline-aware explanation mapping can increase transparency while limiting compute. Reporting of expression complexity, input-space fidelity, and stability under resampling remains inconsistent. In vertical federation, compact representation exchange and selective re-evaluation can reduce rounds and bandwidth, but pipeline-composition leakage and verification cost are rarely measured end-to-end. Differential privacy supports feature selection and explanations, but pipeline-level budget accounting and validation of synthetic-data feature importance remain open challenges. Provenance-rich pipelines and meta-feature profiling support reproducible deployments, yet standardized scorecards and head-to-head benchmarks remain scarce.