Interpreting support vector machines applied in laser-induced breakdown spectroscopy.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Laser-induced breakdown spectroscopy (LIBS) often uses Support Vector Machines (SVMs) for quantitative/qualitative analysis.
• The "black box" nature of SVMs poses risks in high-stakes applications and limits model interpretability.
• Understanding feature importance is crucial for refining SVM models and ensuring reliable results.

Purpose of the Study:

• To develop and compare four novel approaches for interpreting SVMs.
• To investigate feature importance in the classification of 19 algal and cyanobacterial species using LIBS data.
• To assess the impact of feature importance metrics on SVM performance and model bias.

Main Methods:

• Applied four distinct feature importance metrics to SVM models.
• Compared these metrics with established feature selection techniques.
• Analyzed feature weighting in linear and radial basis kernel SVMs.
• Evaluated SVM performance and identified model biases.

Main Results:

• The four feature importance metrics provided complementary and comparable information.
• Identified a bias in SVM models towards features with high variance but significant class overlap.
• Demonstrated that linear and radial basis kernel SVMs assign similar feature weights.
• Feature importance analysis aids in understanding and optimizing SVMs for LIBS applications.

Conclusions:

• Feature importance metrics offer valuable insights into SVM decision-making processes.
• Understanding model bias is critical for improving the reliability of LIBS-based classification.
• This work facilitates more transparent and tunable SVM applications in spectroscopy.