Synergistic eigenanalysis of covariance and Hessian matrices for enhanced binary classification on health datasets

Agus Hartoyo, Jan K. Argasinski, Aleksandra Trenk, Kinga Przybylska, Anna Blasiak, Alessandro Crimi

Abstract: Highlights•Novel data projection method combining covariance and Hessian matrices for enhanced binary classification.•Achieves optimal class separability by leveraging LDA criteria, emphasizing both separation and compactness.•Highlights the importance of ideal dataset conditions, including isotropy and dominant leading eigenvalues.•Empirical validation shows superior performance over traditional dimensionality reduction methods.•Offers insights into deep neural network decision-making processes through eigenanalysis.