Machine Learning Classification of Public Tenders Using PCA: Facilitating SME Access to Chilean Procurement (#2733)

Authors

Caroca, Alejandro

Ruete, David

Varas, Angélica

Marambio-Correa, Francisco

Salinas, Omar

San Martin Medina, Lilian

Maidana, Jean

Abstract

Finding relevant procurement opportunities in Chile's Mercado Público platform is genuinely difficult for small suppliers. The platform processed over 2 million purchase orders in 2024, but its category taxonomy is inconsistently applied, descriptive fields vary in quality, and there is no automated filtering tailored to provider needs. This paper asks whether structured administrative variables alone, without any processing of tender text, can support accurate automatic classification of public tenders into procurement categories. We work with 204,770 tender records from 2024 and focus on the 10 most frequent combinations, covering 56,555 records. Principal Component Analysis (PCA) reduces 14 structured variables to 10 components retaining 89.82% of variance; six classifiers are then trained and compared: Random Forest, XGBoost, Decision Tree, Logistic Regression, SVM, and k-Nearest Neighbors. Random Forest reaches 91.17% accuracy and an F1-macro score of 0.9112. A single Decision Tree comes within 0.06 percentage points of that figure, which has practical implications for deployments where audit trails matter. The full pipeline was implemented independently in both R and Python, with under 1% accuracy difference between environments, confirming reproducibility. The gap between tree-based methods and Logistic Regression exceeds 51 percentage points, confirming that the classification boundaries are non-linear and cannot be captured by simpler models. These results suggest that structured-variable classification is a viable foundation for SME-facing tender recommendation systems in procurement platforms with heterogeneous or incomplete text data.