Seleção de variáveis baseada em algoritmo firefly e análise discriminante por mínimos quadrados parciais para problemas de classificação

Abstract

Variable selection is a potential tool that seeks to filter relevant information to solve problems involving complex matrices. Seeking to improve the accuracy of chemometric models and robustness to the principle of parsimony different algorithms have been developed using variable selection. With the advances in artificial intelligence the use of bio-inspired algorithms for the optimization and resolution of complex problems has become an interesting tool for several applications in the context of multivariate calibration and classification. In this context, the present study proposes a new bio-inspired algorithm in the behavior of fireflies called FA-PLS-DA for variable selection employing Linear Discriminant Analysis seeking to overcome problems involving high multicollinearity among variables. To evaluate the performance of the proposed algorithm, we stipulated population conditions of 50 fireflies in 50 life cycles and used three case studies involving public domain NIR spectrometric databases and a database with simulated information. The first database consists of ATR-FTIR spectra in the 4000 to 650 cm-1 range of 104 saliva samples to assess the presence or absence of SARS-CoV-2. The second database used also consists of NIR spectra of 192 goat milk samples to assess adulteration by the addition of cow's milk. The third database also consists of NIR spectra of 120 samples of extra virgin olive oil from four different countries. For study with simulated information the database comprised 90 observations with 600 variables using four factors to generate three distinct classes, the samples were divided into training and test sets using the Kennard-Stone algorithm. The performance of FA-PLS-DA was compared with the results of Partial Least Squares Linear Discriminant Analysis (PLS-DA) applying different preprocessing to the data, and the data was treated in programs developed in Matlab environment. The preprocessed models that presented the highest TCC were selected, the FA-PLS-DA algorithm selected 3, 7, 9 and 7 latent variables (LVs) for the COVID, goat milk, extra-virgin olive oil and simulated databases respectively, while the PLS-DA selected 4, 16, 16 and 4 LVs. Furthermore, the proposed algorithm showed a correct classification rate (CCR) of 100% for the COVID database training set, while PLS-DA showed a rate of 98.72% by employing a larger number of latent variables. For the goat milk database, the proposed algorithm showed a CCR of 95.92%, while PLS-DA showed 100% CCR. Despite outperforming the proposed algorithm in terms of CCR, PLS-DA employed a high number of LVs for model building. The proposed algorithm outperformed PLS-DA in building the models for the olive oil database, where it obtained 100% CCR for the training and test sets by employing the smallest number of LVs. For the data with simulated information FA-PLS-DA showed 82, 22% CCR for the test set, while PLS-DA showed 77.78%. For all the databases, the FA-PLS-DA algorithm proved to be more parsimonious than PLS-DA, and its performance compared to PLS-DA was robust and able to classify the samples properly from the selected variables, corroborating its viability.