Performance Analysis of Various Supervised Binary Classification Algorithms and their Optimized Variants on High-Dimension Limited-Sample-Size Data

In modern times, high-dimension limited-sample-size (HDLSS) data have become a lot prevalent in data mining research areas. Performing analysis on HDLSS data is inherently difficult because of large number of features and low sample size. Dimensionality reduction is a recommended step before working on HDLSS data because of a phenomenon known as the ‘curse of dimensionality’. However in medical and psychiatric areas, these large number of features often prove to be extremely important in cases of life threatening disease detection or identifying biomarkers. The purpose of this study is to perform a comparative performance analysis of various supervised classification algorithms and their optimized variants on HDLSS data without any dimensionality reduction to recreate the aforementioned situations. Various hyper-parameter optimization techniques and AdaBoost meta-estimator classifier are implemented on these algorithms to find their optimized variants. Optimized Random Forest and CatBoost classifiers produced the most optimal results on the used HDLSS datasets. Relevant arguments and findings regarding the used eight different supervised classification algorithms, both base and optimized, are evaluated and summarized in this paper.