Uncertainty estimation and misclassification probability for classification models based on discriminant analysis and...

Area of Science:

• Machine Learning
• Chemometrics
• Data Science

Background:

• Uncertainty estimation quantifies classification model predictive performance via misclassification probability.
• Low misclassification probability signifies high model trustworthiness, while high probability indicates susceptibility to errors.

Purpose of the Study:

• To develop misclassification probability estimations for classification models using uncertainty estimation.
• To evaluate the robustness and reliability of these estimations across different datasets and noise conditions.

Main Methods:

• Utilized bootstrap-based uncertainty estimation for discriminant analysis (LDA, QDA) and support vector machines (SVM).
• Employed Principal Component Analysis (PCA) for variable reduction prior to classification.
• Tested models on one simulated and three real-world spectral datasets for binary and ternary classifications.

Main Results:

• Models with lower misclassification probabilities demonstrated greater stability under white Gaussian noise perturbation, indicating enhanced robustness.
• Misclassification probability served as a reliable metric for assessing classifier robustness.

Conclusions:

• Misclassification probability is a valuable metric for evaluating classifier robustness and predictive performance.
• Uncertainty estimation provides a quantitative measure of model trustworthiness in classification tasks.