Interpreting support vector machines applied in laser-induced breakdown spectroscopy

Interpreting support vector machines applied in laser-induced breakdown spectroscopy

Článek WoS

Laser-induced breakdown spectroscopy is often combined with a multivariate black box model—such as support vector machines (SVMs)—to obtain desirable quantitative or qualitative results. This approach carries obvious risks when practiced in high-stakes applications. Moreover, the lack of understanding of a black-box model limits the user's ability to fine-tune the model. Thus, here we present four approaches to interpret SVMs through investigating which features the models consider important in the classification task of 19 algal and cyanobacterial species. The four feature importance metrics are compared with popular approaches to feature selection for optimal SVM performance. We report that the distinct feature importance metrics yield complementary and often comparable information. In addition, we identify our SVM model's bias towards features with a large variance, even though these features exhibit a significant overlap between classes. We also show that the linear and radial basis kernel SVMs weight the same features to the same degree.

Laser-induced breakdown spectroscopy is often combined with a multivariate black box model—such as support vector machines (SVMs)—to obtain desirable quantitative or qualitative results. This approach carries obvious risks when practiced in high-stakes applications. Moreover, the lack of understanding of a black-box model limits the user's ability to fine-tune the model. Thus, here we present four approaches to interpret SVMs through investigating which features the models consider important in the classification task of 19 algal and cyanobacterial species. The four feature importance metrics are compared with popular approaches to feature selection for optimal SVM performance. We report that the distinct feature importance metrics yield complementary and often comparable information. In addition, we identify our SVM model's bias towards features with a large variance, even though these features exhibit a significant overlap between classes. We also show that the linear and radial basis kernel SVMs weight the same features to the same degree.

LIBS, Classification, Feature importance, SVM, Interpretable machine learning

LIBS, Classification, Feature importance, SVM, Interpretable machine learning

KÉPEŠ, E.; VRÁBEL, J.; ADAMOVSKÝ, O.; STŘÍTEŽSKÁ, S.; MODLITBOVÁ, P.; POŘÍZKA, P.; KAISER, J.

2023

01.02.2022

1873-4324

Analytica Chimica Acta

1192

339352

Nizozemsko

1

12

12

https://www.sciencedirect.com/science/article/pii/S0003267021011788#appsec1