Abordagem Não-Linear de Redução de Dimensionalidade em Duas Etapas aplicada à classificação de engajamento em Redes Sociais

Abstract

This thesis presents a data classification methodology entitled NL-TSDR (Nonlinear Two-Stage Dimensionality Reduction), aimed at mitigating the challenges posed by high- dimensional datasets and their inherent class masking problems. This phenomenon occurs when overlapping instances from di!erent groups lead to the degradation of traditional classification methods’ performance. The proposed method comprises two distinct stages. In the first stage, the objective is to extract discriminative features by maximizing pairwise class separation probability, e!ectively mitigating class overlap. The second stage applies nonlinear transformations (hyperbolic tangent, Gaussian, sigmoid, and Laplacian functions) to enhance class center separability through multi-objective optimization that maximizes inter-class distances while minimizing intra-class dispersion. Experimental validation was conducted on benchmark datasets and real social media engagement data. The method demonstrates consistent improvements across all evaluated scenarios, achieving superior accuracies even with significant dimensional reduction. Specifically, on the social media dataset, the method achieves a 3.04% accuracy improvement over baseline methods, simultaneously obtaining a dataset that is five and a half times smaller than the original. The analysis of variance (ANOVA) with p = 0.0034, corroborated by paired t-test (p = 0.0199) and Wilcoxon test (p = 0.0153), indicates a statistically significant di!erence (p → 0.05) between the classification accuracies of the original data and the data processed by the proposed method. The results demonstrate that NL-TSDR provides an e!ective tool for addressing class masking in high-dimensional classification tasks while maintaining computational e"ciency suitable for real-world applications.