Exploring Variable Selection Methods and Data Fusion in Support Vector Regression: An Application in Petroleomics
Name: PEDRO HENRIQUE PEREIRA DA CUNHA
Publication date: 28/03/2024
Examining board:
Name |
Role |
|---|---|
| LUCAS MATTOS DUARTE | Examinador Externo |
| MARIANA RAMOS DE ALMEIDA | Examinador Externo |
| MURILO DE OLIVEIRA SOUZA | Examinador Externo |
| PAULO ROBERTO FILGUEIRAS | Presidente |
| WANDERSON ROMÃO | Examinador Interno |
Summary: Support Vector Regression (SVR) is considered a black-box machine learning method and has stood out in chemometrics over the past decades, achieving results superior or equal to methods already established in academia. As a black-box method, it is challenging to understand the cause/effect relationship. To address this, variable selection can be applied, a strategy that aims to identify the most influential variables in building the model. This work proposes the development of two variable selection methods - Permutation Subwindow Analysis (SPA) and Noise-Incorporated Permutation Subwindow Analysis (NISPA) - to apply in SVR combined with infrared. SPA and NISPA provided the most accurate models for kinematic viscosity, saturates, and aromatic content. The root mean square error of prediction (RMSEP) for SPA and NISPA were, respectively, 14.3% and 14.6% for kinematic viscosity, 4.7% and 4.4% for saturates content, and 3.4% and 3.1% for aromatic content. Therefore, SPA and NISPA, in addition to generally obtaining faster, more accurate, and more parsimonious models, revealed the most important variables for building SVR models. Another way to improve a model is data fusion, but this strategy has been little studied in SVR. Thus, data fusion was studied using NIR, MIR, and NMR of ¹H and ¹³C combined using low, medium, and high-level fusion. The models generated by data fusion were superior to the models without fusion for most tests. In API density, the application of medium-level fusion using PCA combining MIR and NIR developed a model with better parameters than the model without data fusion. By applying mediumlevel fusion with GA to predict pour point, combining NIR and NMR of ¹H, it was possible to surpass models without fusion, as well as models found in the literature. In total nitrogen, high-level fusion with MIR and NMR of ¹H proved to be statistically better than models without data fusion. This demonstrates that it is possible to extract new information for SVR modeling using data fusion and obtain statistically better models than those derived from isolated analytical sources.
