Explainable model of supervised and unsupervised learning to characterize the coverage of social programs in vulnerable territories of Peru

Marin-Rodriguez W.J.; Vellón-Flores V.I.; Ausejo-Sánchez J.L.; Solano-Armas T.; Maguiña-Poma Y.E.; Samanamud-Malca S. · Revista Cientifica de Sistemas e Informatica (2026) · DOI: 10.51252/rcsi.v6i2.1266

Resumen

This study developed and evaluated an explainable supervised and unsupervised learning approach to characterize social program coverage and identify historically vulnerable territories in Peru. A total of 35,952 district-month observations from 1,893 districts and 19 periods were integrated using MIDIS administrative data and poverty maps. Based on 13 coverage indicators, four clustering algorithms and three classifiers were compared through cross-validation and an independent test set. K-means with two clusters was selected as the main solution, achieving a silhouette score of 0.540 and identifying a profile associated with greater coverage and historical vulnerability. For classification, Elastic Net using only coverage indicators achieved an F1-score of 0.655 and, in the independent test set, a sensitivity of 0.829, an F1-score of 0.647, and a ROC-AUC of 0.871, with no significant differences compared with the other models. Explainability analysis highlighted Juntos coverage, estimated FONCODES users, and the number of programs present. The findings show that administrative records can support territorial prioritization when complemented with updated information and institutional review. © Los autores.

Instituciones: Universidad Nacional José Faustino Sánchez Carrión