Evaluating Dimensionality Reduction Methods; A Focus on linear Data

The work explores the impact of dimensionality reduction techniques on supervised machine learning classifiers, focusing specifically on socio-economic and developmental data from South Asian countries. A very large number of variables in a dataset often complicates the analysis process and diminishes model efficiency. To overcome this challenge, the classification accuracy of "Various dimensionality reduction techniques, such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Kernel Principal Component Analysis (KPCA), Linear Discriminant Analysis (LDA), Kernel PCA (KPCA), and T-Distributed Stochastic Neighbor Embedding (t-SNE), were evaluated. The results suggest that the classification accuracy is best in case of LDA among the other techniques. In addition, LDA-based socioeconomic categorization for South Asia is highly suitable when the data sets are class well defined. The PCA was balanced, but somewhat average, and showed value for reducing noise and simplifying a complicated data set. KPCA was focused on uncovering a relationship that was not linear and only resulted in small gains for the present data set. This paper does reflect an important aspect to choose an appropriate strategy to align dimension reduction with the inherent data structure and the task for the analysis. The insights presented here are designed to guide researchers and policymakers in South Asia to select appropriate techniques for maximizing classification accuracy, interpretability, and computational efficiency. These findings advance data-driven decision making and machine learning applications in socio-economic research.