ENHANCED INTERPRETABILITY USING UMAP-BASED LOCAL BIPLOTS FOR THE ANALYSIS OF COFFEE PRODUCTION IN HUILA

Abstract

This study applies an interpretable machine learning framework based on UMAP dimensionality reduction and visualization through Local Biplots to analyze the structural complexity of specialty coffee production in the department of Huila, Colombia. Using a multivariate database covering the period 2001–2025 and integrating productive, climatic, agronomic, and economic variables, the methodology identifies latent patterns, heterogeneous production regimes, and nonlinear relationships that remain hidden under conventional linear approaches. The UMAP-based Local Biplot reveals compact and coherent clusters that capture local structures in the reduced space, outperforming the classical SVD-Biplot, which exhibits overlap and limited separation among observations. The results show that the relevance of key variables varies substantially across clusters, highlighting the influence of seasonality, solar radiation, production costs, and labor intensity in shaping local production dynamics. Likewise, the Kernel Density Estimation (KDE) applied to temporal segments evidence transitions from concentrated structures in the early periods to more dispersed and diverse configurations in recent years, likely associated with climatic variability, technological adoption, and crop renovation cycles. Overall, the findings demonstrate that nonlinear and interpretable methods provide a deeper understanding of the multidimensional factors that determine coffee productivity, surpassing the limitations of aggregated indicators and linear techniques.