AN EXPLAINABLE AI FRAMEWORK FOR TRANSPARENT POVERTY CLASSIFICATION IN NIGERIA

ABSTRACT

The ineffectiveness of poverty targeting mechanisms in Nigeria is reflected in the exclusion errors of over 40% and the fact that Nigeria's roughly 6.8 million households have no access to welfare services. The existing Proxy Means Test (PMT) systems are opaque, unclear and binary, as provided for in the Nigeria Data Protection Act of 2023. This paper presents an innovative method for explainable artificial intelligence (XAI) using Fuzzy-Adaptive Stacking Ensemble (FAS-XAI) combining Type-2 Fuzzy Logic, ensemble learning approaches (XGBoost, CatBoost, LightGBM) and a transparency module. The proposed model has achieved a high predictive power (R² = 0.967; AUC = 0.996) and reduced the exclusion error to 34.3% using the Nigeria General House Survey (GHS) Panel Wave 5 dataset (2023–2024; n = 5,067). The Fuzzy transformation of volatile variables is credited with this improvement, as it helps to make the learning paradigm more stable. The simulations also demonstrate that multi-domain policy interventions achieve a poverty reduction rate of 31.2%, a higher rate than single-domain policy interventions (28.8%). In this investigation, we offer a framework for uncertainty-aware welfare targeting, which is transparent, to support the achievement of the Sustainable Development Goals (SDGs) 1, 5, 10, and 16.

Keywords: Multidimensional poverty, Type-2 Fuzzy Logic, stacking ensemble, explainable AI,