Interpretable Learning for Identifying Digital Access Vulnerability in Underserved Communities

Unequal access to digital technologies results from interacting socioeconomic, demographic, geographic, behavioral, and infrastructural conditions that cannot be represented adequately through broadband availability alone. This study proposes an interpretable analytical framework for identifying digital access vulnerability and prioritizing underserved communities without requiring proprietary individual level information. The framework integrates composite index construction, sparse classification, monotonic predictive modeling, probability calibration, clustering, and fairness evaluation. A standardized feature vector, zi represents internet affordability, device ownership, unemployment, educational participation, population vulnerability, service coverage, and digital service utilization. The Digital Access Vulnerability Score (DAVS) summarizes multidimensional vulnerability, while the predictive model estimates the probability of high digital-access vulnerability. Community profiling identifies infrastructure-limited, affordability limited, device-limited, and multidimensionally underserved groups. SHAP values, partial dependence analysis, sparse coefficients, and counterfactual explanations provide transparent interpretation of prediction outcomes. The framework uses simulated county or community level data, with optional validation using public broadband, census, housing, employment, and digital adoption datasets. The proposed framework provides a transparent and reproducible decision-support approach for identifying vulnerability, prioritizing interventions, and evaluating prediction reliability and fairness.