Dimensionality Reduction and Clustering of Global Large Enterprise Data Using PCA, UMAP, and Gaussian Mixture Models

Abstract

In the modern business landscape, large corporations generate high dimensional datasets that combine financial, operational, and market indicators, often producing complex and partially overlapping structures that are difficult to interpret in the original feature space. This study benchmarks linear dimensionality reduction using Principal Component Analysis (PCA) against non linear reduction using Uniform Manifold Approximation and Projection (UMAP), and examines how these representations affect clustering quality using k means and Gaussian Mixture Models (GMM). Data preprocessing includes missing value handling, categorical encoding, numeric coercion, and feature standardization to ensure scale comparable learning. Clustering performance is evaluated using the silhouette score, which jointly reflects within cluster cohesion and between cluster separation. The results indicate that the UMAP plus GMM pipeline achieves the best clustering quality with the highest silhouette score (0.57), suggesting that manifold based representations combined with probabilistic clustering more effectively capture heterogeneous corporate structures than linear projections and hard assignments. The findings support the use of non linear and model based pipelines for corporate segmentation tasks, particularly when clusters may overlap due to mixed business profiles and cross sector similarities.