Two-Phase Dimensionality Reduction Using Representative Selection and UMAP Training

Abstract

High-dimensional data is increasingly prevalent across diverse domains such as bioinfor matics, medical imaging, and natural language processing, posing significant challenges due to the curse of dimensionality and computational complexity. This thesis proposes a novel two-phase dimensionality reduction framework that combines representative selec tion through clustering with Uniform Manifold Approximation and Projection (UMAP) training. In the first phase, representative samples are selected using clustering algorithms such as Mini-Batch KMeans and BIRCH to reduce data size while preserving its structure. In the second phase, UMAP is trained on these representatives to learn a low-dimensional embedding, which is then used to transform the entire dataset efficiently. Experimental results on the IoTID20 dataset; a high-dimensional dataset, demonstrate that the pro posed method significantly reduces computational time and memory usage compared to standard UMAP, while maintaining comparable embedding quality and classification per formance. This hybrid approach offers a scalable and effective solution for dimensionality reduction in large-scale high-dimensional data analysis