Explainable AI Maps Broadband Gaps, Guides Investment Strategy

The United States has allocated substantial funds for broadband expansion, yet effective, evidence-based strategies for targeting these investments remain underdeveloped. This paper introduces an explainable machine learning framework designed to precisely profile broadband adoption disparities at the census-tract level across over 83,000 tracts nationwide. The model leverages 65 socioeconomic, demographic, and infrastructure features from the American Community Survey 2022. A LightGBM model, trained with spatial five-fold cross-validation, achieved strong predictive performance, confirmed by state-held-out cross-validation, demonstrating its generalization capability. Crucially, TreeSHAP analysis was employed to identify the most influential factors. Income and education emerged as the dominant group, with an engineered interaction term absorbing significant attribution. SHAP-based clustering further revealed three distinct factor profiles: "Well-Connected Moderate," "Affordability-Limited Severe," and "Rural-Elderly." As a screening tool, the ML-based tract selection captured a larger portion of the total adoption gap within the top 10% of tracts compared to income-only heuristics. The primary contribution lies in the per-tract factor decomposition, where SHAP identifies which feature groups are most strongly associated with each tract's predicted gap, enabling differentiated investigation and more informed policy decisions. The model also demonstrated temporal stability, maintaining ranking accuracy when trained on older data and predicting newer outcomes.