Fairness-Aware Prediction of Digital Financial Inclusion Risk in Informal Paratransit: A Case Study from Ibadan, Nigeria

Digital fare-payment systems are expanding in Nigeria's public road transport sector, while informal paratransit remains predominantly cash-based. In Ibadan (Oyo State, population ~4.3 million), the state-owned Pacesetter Transport Services operates a digital-payment system alongside informal modes, minibuses (Korope), taxis (Micra), tricycles (Keke NAPEP), and motorcycles (Okada). As predictive models increasingly identify users at risk of digital exclusion, models trained on formal-system data risk failing to capture informal-paratransit users' socioeconomic and digital characteristics, producing unequal prediction errors across income, education, and digital exposure precisely among the riders inclusive digitalization is meant to reach. We develop a fairness-aware framework for predicting digital financial inclusion risk among Ibadan's informal-paratransit users. Data (collected). Primary survey data from 300 respondents, 200 informal operators (120 Micra, 80 Keke) and 100 passengers, stratified across five transport hubs (Mokola, Moniya, Bodija, Gate, Ojoo), capturing income, education, banking access, digital-payment confidence, and Pacesetter exposure. Digital financial inclusion risk is defined as the likelihood of difficulty participating in digitally-mediated transport payment and related financial services. Analysis (completed). Descriptive statistics confirm the disparities motivating a fairness framing: all passengers report holding a bank account, yet digital-payment confidence skews sharply toward already-comfortable users (72% "very comfortable" vs. 7% neutral), and Pacesetter exposure is unevenly distributed by hub (30% "occasionally," 5% "regularly," 38% report no route access), suggesting infrastructure exposure, not individual readiness, drives adoption gaps. Modeling (in progress). We frame inclusion-risk prediction as binary classification using income, education, banking access, digital-payment confidence, hub location, and Pacesetter exposure as features. Planned models: logistic regression (interpretable baseline) and XGBoost (nonlinear comparator), benchmarked against fairness-constrained variants using Fairlearn's exponentiated-gradient reduction with equalized odds imposed across income-bracket and hub-defined subgroups. Evaluation: AUC and accuracy for performance; equalized-odds difference, demographic-parity difference, and subgroup calibration for fairness. Hypothesis: unconstrained models will underpredict inclusion risk for hub-peripheral, lower-income subgroups relative to constrained variants, with minimal accuracy trade-off. This work extends fairness-aware machine learning to a population structurally underrepresented in formal transport datasets, demonstrating how equity constraints can be built into inclusion-risk models directly rather than audited post-deployment. It provides an empirical basis for more inclusive digital mobility and financial-service systems in African cities. Keywords: Fairness-aware machine learning; digital financial inclusion; informal paratransit; algorithmic fairness; digital mobility; Ibadan, Nigeria. References Zheng, Wang & Zhao (2022). Transportation Research Part C, 132, 103410. Bird et al. (2020). Fairlearn: A toolkit for assessing and improving fairness in AI. Microsoft Research Technical Report. Osabutey & Jackson (2024). Technological Forecasting & Social Change, 202, 123339.