水中有机污染物与臭氧反应速率常数的可解释性机器学习模型  

An Explainable Machine Learning Model for Reaction Rate Constantsof Organic Compounds in Water with Ozone

作　　者：刘洋[1] 魏宠芝 孙婷 任月英 LIU Yang;WEI Chong-zhi;SUN Ting;REN Yue-ying(School of Environmental and Municipal Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学环境与市政工程学院,甘肃兰州730070

出　　处：《兰州文理学院学报（自然科学版）》2024年第6期91-106,共16页Journal of Lanzhou University of Arts and Science(Natural Sciences)

摘　　要：运用多元线性回归(MLR)、非线性支持向量回归(SVR)以及投影寻踪回归(PPR)方法构建了定量结构-活性关系(QSAR)模型,以预测水相有机物与臭氧的反应速率常数(logK_(O_(3))).采用内检验和外检验方法,对模型的拟合性能、稳健性以及预测能力进行了比较.结果表明,非线性模型结果优于线性模型;PPR模型的性能最佳.采用SHAP方法对PPR进行可视化表征及分析,以提高模型预测结果的可靠性,增强模型的透明度,从而弥补了机器学习模型的“黑箱”缺陷.最后,利用Williams图法表征了PPR模型的应用域范围(AD).By using multiple linear regression(MLR),nonlinear support vector regression(SVR)and projection pursuit regression(PPR)methods,QSAR models were developed to predict the reaction rate constant(log K O 3)of organic compounds with ozone in water.The robustness,fitting performance and ability of the models were compared by internal and external validation procedures.The results showed that,comparatively,nonlinear models were better than linear model,with PPR performs best.Then the PPR model was visualized and explained through SHAP analysis,providing interpretability and explainability for the black-box nature of machine learning model,to strengthen the transparency and credibility of the model.In addition,the applicability domain(AD)range of PPR models was defined and visualized via a Williams plot.

关 键 词：水环境 定量结构-活性关系(QSAR) 有机污染物 臭氧反应速率 投影寻踪回归 SHAP 

分 类 号：X-131[环境科学与工程]