An Explainable Ai Framework For Transparent Poverty Classification And Citizen Engagement In Nigeria

Abstract

Poverty targeting in Nigeria remains quite astonishingly inefficient with exclusion error rates above 40 per cent which puts millions of eligible households out of reach of welfare assistance. The existing models of the Proxy Means Test (PMT) are binary classification based, non-transparent and cannot work in dynamic and high noise settings and this leads to an ongoing accuracy-transparency-robustness trilemma. The aim is to design and test an explainable artificial intelligence system that will increase the accuracy of poverty classification, transparency, and decrease errors of exclusion in the welfare targeting system of Nigeria. The Design Science Research (DSR) methodology was applied to develop the Fuzzy-Adaptive Stacking Ensemble for Explainable AI (FAS-XAI) that incorporates Type-2 Fuzzy Logic, stacking ensembles of XGBoost, CatBoost, and LightGBM, and a Cognitive Transparency Module. This model was evaluated using the GHs Wave 5 (20232024; N = 5,067) of Nigeria with cross-validation and performance values of R 2 and AUC. FAS-XAI showed an impressive predictive performance (R 2 = 0.967; AUC = 0.996), reducing the exclusion errors by 100-34.3 per cent. High-ranked predictors were found to be the dependency ratio, asset wealth and gaps in energy transition, whereas integrated interventions had more significant poverty reduction impacts. This paper introduces a novel groundbreaking fuzzy-stacking explainable AI framework that combines interpretability and robustness, providing a policy-relevant, scalable solution to transparent and equitable poverty targeting in Nigeria.