Scholarlink

<p>Quantitative analysis of high-dimensional spectral data presents significant computational challenges, characterized by a combinatorial explosion of potential feature subsets and hyperparameter configurations. Navigating this vast search space to construct robust models requires advanced feature selection and rigorous algorithmic tuning—processes often computationally prohibitive for manual experimentation and prone to overfitting. To address these issues, we introduce AutoRegress, a high-throughput computing (HTC)-enabled automated machine learning (AutoML) framework designed to systematize and accelerate regression modeling for spectral datasets. AutoRegress leverages parallel processing to execute a computationally intensive hybrid nested cross-validation protocol, ensuring statistical rigor and mitigating the optimistic bias common in single-split analyses. The framework orchestrates a multi-stage Bayesian optimization pipeline that dynamically selects between linear—partial least squares (PLS), Ridge, and Elastic Net—and nonlinear support vector regression algorithms while encapsulating filter-based feature selection techniques within strictly isolated training folds to prevent data leakage. The efficacy of the framework was validated through two distinct case studies. In the first, predicting octane numbers from standard near-infrared spectra, AutoRegress identified a Ridge regression model paired with a Pearson correlation filter, reducing the feature space by 70%. This configuration achieved an external test <i>R</i><sup>2</sup> of 0.9722 and a root mean square error of prediction of 0.2256, yielding results competitive with complex neural network benchmarks. The second case study involved the simultaneous quantification of pharmaceutical analytes, itraconazole (ITZ) and secnidazole (SEZ). For ITZ, the framework leveraged the intrinsic regularization of Elastic Net on the full spectrum (<i>R</i><sup>2</sup> = 0.9960), while for SEZ, it identified PLS with Analysis of Variance (ANOVA)-based feature selection as the optimal strategy (<i>R</i><sup>2</sup> = 0.9994). By automating the evaluation of over 2800 distinct models, AutoRegress demonstrates how HTC architectures can be utilized to standardize advanced chemometrics, facilitating reproducible, high-precision analytical workflows with minimal user intervention.</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

AutoRegress: a user-friendly Bayesian-optimized framework for robust regression modeling and feature selection in spectral analysis

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Afnan Altwala,
  • Ayman M. Algohary,
  • Ahmed M. Ibrahim

摘要

Quantitative analysis of high-dimensional spectral data presents significant computational challenges, characterized by a combinatorial explosion of potential feature subsets and hyperparameter configurations. Navigating this vast search space to construct robust models requires advanced feature selection and rigorous algorithmic tuning—processes often computationally prohibitive for manual experimentation and prone to overfitting. To address these issues, we introduce AutoRegress, a high-throughput computing (HTC)-enabled automated machine learning (AutoML) framework designed to systematize and accelerate regression modeling for spectral datasets. AutoRegress leverages parallel processing to execute a computationally intensive hybrid nested cross-validation protocol, ensuring statistical rigor and mitigating the optimistic bias common in single-split analyses. The framework orchestrates a multi-stage Bayesian optimization pipeline that dynamically selects between linear—partial least squares (PLS), Ridge, and Elastic Net—and nonlinear support vector regression algorithms while encapsulating filter-based feature selection techniques within strictly isolated training folds to prevent data leakage. The efficacy of the framework was validated through two distinct case studies. In the first, predicting octane numbers from standard near-infrared spectra, AutoRegress identified a Ridge regression model paired with a Pearson correlation filter, reducing the feature space by 70%. This configuration achieved an external test R2 of 0.9722 and a root mean square error of prediction of 0.2256, yielding results competitive with complex neural network benchmarks. The second case study involved the simultaneous quantification of pharmaceutical analytes, itraconazole (ITZ) and secnidazole (SEZ). For ITZ, the framework leveraged the intrinsic regularization of Elastic Net on the full spectrum (R2 = 0.9960), while for SEZ, it identified PLS with Analysis of Variance (ANOVA)-based feature selection as the optimal strategy (R2 = 0.9994). By automating the evaluation of over 2800 distinct models, AutoRegress demonstrates how HTC architectures can be utilized to standardize advanced chemometrics, facilitating reproducible, high-precision analytical workflows with minimal user intervention.