Counterfactual explanation-based targeted regulation of key chemical indicators for improving tobacco leaf sensory quality
- 1Zhengzhou Tobacco Research Institute of CNTC
- 2Technology R&D Center, Gansu Tobacco Industrial Co., Ltd.
- 3Technology Center, China Tobacco Shaanxi Industrial Co., Ltd.
Abstract
Introduction: To address the need for the value-added utilization of tobacco leaves with low sensory quality, this study proposed a targeted chemical composition regulation method based on counterfactual explanations and preliminarily evaluated the resulting chemical adjustment strategies using samples processed by a rotary-drum tobacco high-temperature (THT) drying process.
Methods: Based on 70 chemical indicators and sensory smoking evaluation data from 354 tobacco leaf samples, binary shrinkage linear discriminant analysis (shrinkage LDA) models were developed to distinguish between the high- and low-sensory-quality groups in terms of aroma quality, aroma intensity, and offensive odor. Incremental feature selection (IFS) was employed to identify key chemical indicators, after which indicators with low adjustability under THT conditions were further excluded. The decision models for counterfactual explanations were then reconstructed using the remaining indicators. Sequential least squares programming (SLSQP) was applied to generate chemical-indicator adjustment pathways for improving sensory quality, and the resulting pathways were validated through a THT experiment conducted at 215 °C.
Results: The results showed that: (1) following incremental feature selection and further screening based on the adjustability of chemical indicators under THT conditions, 11, 9, and 10 key chemical indicators were ultimately retained in the decision models for aroma quality, aroma intensity, and offensive odor, respectively; (2) all the developed models achieved classification accuracies above 84% on the test sets, indicating good mapping relationships between the key chemical indicators and sensory quality; (3) the directional agreement between the theoretically predicted adjustments and the experimentally measured changes ranged from 66.7% to 72.7%. Moreover, indicators showing directional agreement contributed significantly more to improvements in discriminant scores than those showing opposite directions, confirming their dominant role in sensory quality improvement.
Discussion: These findings demonstrate that counterfactual explanations can effectively predict optimization pathways in complex chemical systems, thereby providing theoretical support and methodological guidance for the targeted improvement of tobacco leaves with low sensory quality.
Related
- Stress-Guided Localized Infill Variation for Improving the Mechanical Efficiency of FDM-Printed L-Bracket Structures
- Open Ontology of Polymer Chemistry: Accumulation and Self-Improving Memory Mechanisms for Polymer Research
- A Natural Approach to Improving Berry Quality in Table Grapes: Inactivated Yeast Extracts
- Evaluation of intelligent aerogel indicators composed of red cabbage extract and pectin for monitoring chicken meat freshness.
- Poly(lactic acid)/ZIF-67 colorimetric headspace indicators for monitoring chicken breast freshness.
BibTeX
@article{chen2026counterfactual,
title = {{Counterfactual explanation-based targeted regulation of key chemical indicators for improving tobacco leaf sensory quality}},
author = {Hexin Chen and Wenting Li and Junwei Guo and Hongbo Wang and Song Yang and Ran Wan and Yanling Liu and Zhenyu Liu and Guoxin Qin and Cong Wang and Zeli Wang and Haiyue Su and Xianyue Zhu and Zhen Wang and Zhicai Wang and Cong Nie and Le Zhao and Shan Li and Xufeng Wang},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1991648},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1991648/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Counterfactual explanation-based targeted regulation of key chemical indicators for improving tobacco leaf sensory quality
AU - Hexin Chen
AU - Wenting Li
AU - Junwei Guo
AU - Hongbo Wang
AU - Song Yang
AU - Ran Wan
AU - Yanling Liu
AU - Zhenyu Liu
AU - Guoxin Qin
AU - Cong Wang
AU - Zeli Wang
AU - Haiyue Su
AU - Xianyue Zhu
AU - Zhen Wang
AU - Zhicai Wang
AU - Cong Nie
AU - Le Zhao
AU - Shan Li
AU - Xufeng Wang
PY - 2026
DA - 2026/10/22
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1991648
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1991648/full
AB - Introduction: To address the need for the value-added utilization of tobacco leaves with low sensory quality, this study proposed a targeted chemical composition regulation method based on counterfactual explanations and preliminarily evaluated the resulting chemical adjustment strategies using samples processed by a rotary-drum tobacco high-temperature (THT) drying process. Methods: Based on 70 chemical indicators and sensory smoking evaluation data from 354 tobacco leaf samples, binary shrinkage linear discriminant analysis (shrinkage LDA) models were developed to distinguish between the high- and low-sensory-quality groups in terms of aroma quality, aroma intensity, and offensive odor. Incremental feature selection (IFS) was employed to identify key chemical indicators, after which indicators with low adjustability under THT conditions were further excluded. The decision models for counterfactual explanations were then reconstructed using the remaining indicators. Sequential least squares programming (SLSQP) was applied to generate chemical-indicator adjustment pathways for improving sensory quality, and the resulting pathways were validated through a THT experiment conducted at 215 °C. Results: The results showed that: (1) following incremental feature selection and further screening based on the adjustability of chemical indicators under THT conditions, 11, 9, and 10 key chemical indicators were ultimately retained in the decision models for aroma quality, aroma intensity, and offensive odor, respectively; (2) all the developed models achieved classification accuracies above 84% on the test sets, indicating good mapping relationships between the key chemical indicators and sensory quality; (3) the directional agreement between the theoretically predicted adjustments and the experimentally measured changes ranged from 66.7% to 72.7%. Moreover, indicators showing directional agreement contributed significantly more to improvements in discriminant scores than those showing opposite directions, confirming their dominant role in sensory quality improvement. Discussion: These findings demonstrate that counterfactual explanations can effectively predict optimization pathways in complex chemical systems, thereby providing theoretical support and methodological guidance for the targeted improvement of tobacco leaves with low sensory quality.
SN - 1664-462X
ER -