选择(遗传算法)
水准点(测量)
统计
回归
树(集合论)
数学
计算机科学
选型
约束(计算机辅助设计)
高斯过程
作文(语言)
组分(热力学)
过程(计算)
数据挖掘
回归分析
机器学习
人工智能
特征选择
生物技术
优先次序
决策树
多元自适应回归样条
二次规划
整数规划
出处
期刊:Molecules
[Multidisciplinary Digital Publishing Institute]
日期:2026-11-16
卷期号:31 (22): 4135-4135
标识
DOI:10.3390/molecules31224135
摘要
Stricter composition screening can prevent unsuitable extraction recommendations while withholding suitable ones. This retrospective benchmark evaluated target compound content and relative composition as decision criteria, comparing selection by predicted target content alone with additional predicted-share screening, using identical predictions and held-out candidates within datasets. The collection comprised 39 datasets from 28 studies, with extensions analyzed separately. Success required measured target content and share within the analytical panel to meet training-derived thresholds. Across seven primary fruit by-product datasets and five focal models, screening increased attainment in 3 of 35 comparisons, by at most 1.4 percentage points, under the original fitting settings, common evaluation sets, and 60th-percentile content and median share thresholds. In grape pomace, reduced quadratic regression gained eight successes but lost 60, including 45 through additional abstention. Saffron floral residues and sage herbal dust provided separate medicinal plant evidence with source-specific limitations. Matched refits and threshold analyses showed that recommendations and some rule comparisons depended on the settings. Assessing compliance with ingredient specifications requires prospective validation across independent material batches using measurements on the relevant product basis. The benchmark supports choosing selection rules by jointly evaluating attainment, unsuitable recommendations, and abstention under specified analytical panels, chemical requirements, and cost assumptions.
科研通智能强力驱动
Strongly Powered by AbleSci AI