A generation of synthetic samples and artificial outliers via principal component analysis and evaluation of predictive capability in binary classification models

Unbalanced sample groups tend to yield models with a higher prevalence of predominant classes. A sample group with balanced classes contributes to the development of more robust models with improved predictive capability to classify classes equally. In the literature, two methodologies for sample balancing can be found: elimination (undersampling) and synthetic sample generation (oversampling). Undersampling methodologies result in the loss of real samples, while oversampling methods may introduce issues related to adding non-real signals to the original spectra. To overcome these challenges, this paper aimed to utilize Principal Component Analysis (PCA) for the generation of virtual samples (synthetic samples and artificial outliers) to balance data in multivariate classification models. The proposed methodology was applied to data from mid-infrared spectroscopy (MIR) and high-resolution mass spectrometry (HRMS) with Partial Least Squares Discriminant Analysis (PLS-DA) and Support Vector Machine (SVM) models. The constructed models demonstrate that the addition of virtual samples enhances performance parameters (e.g., false negative rate, false positive rate, accuracy, sensitivity, specificity, among others) compared to unbalanced models, while also mitigating overfitting (a problem found in unbalanced models). Performance parameters exhibited a more significant improvement percentage using the non-linear model (SVM) compared to the linear model (PLS-DA). Furthermore, the created virtual spectra do not introduce new signals, i.e., original, and virtual spectra exhibit a similar spectral profile, differing only in the intensity levels. Finally, all models demonstrated good predictive capability according to permutation testing for the binary model developed in this work, limiting the rate of class permutation retention (between 40 % and 60 % of the y-vector remained in the original class). All created models exhibited accuracy values higher than the accuracy distribution of models with permuted classes for the test group.