Feature Extraction 및 Random Forest 기법을 이용한 진동분광학 기반 판별 및 정량 분석

Abstract

This thesis presents the non-parametric algorithms as potential quantitative and qualitative chemometric methods for analysis vibrational spectroscopic data. Parametric algorithms such as principal component analysis (PCA) and partial least squares (PLS) are the most widely adopted methods in multivariate analysis; while, possible overfitting and blackbox-like modelling are their drawbacks. On the other hand, non-parametric algorithms could substantially lessen the risk of overfitting and their algorithm structures are simpler to understand. Non-parametric chemometric algorithms including random forest (RF), hit quality index (HQI) and neighborhood preserving embedding (NPE) were employed in this thesis and their analytical performances were evaluated for analysis of sample in diverse fields: 1) quantitative compositional analysis of naphtha and determination research octane numbers (RONs) of gasoline samples using RF, 2) geographical discrimination of agricultural products as well as diesel/light gas oil (LGO) samples using HQI-voting and 3) infrared (IR) spectroscopic diagnosis of stomach and colon malignancy using neighborhood preserving embedding (NPE). In all cases, the resulting accuracies obtained from non-parametric and parametric methods were compared with each other.