A graph neural network approach to physical distress screening in subsidized housing

Abstract

Housing agencies must prioritize physical inspections across tens of thousands of subsidized properties while lacking systematic condition data for most. We demonstrate that Graph Attention Networks (GATs) can propagate distress signals from inspected to uninspected properties through ownership and spatial networks, substantially improving screening efficiency over conventional approaches. Ownership connections add predictive value beyond geographic proximity. Removing them lowers performance, and properties linked to owners with distressed properties elsewhere face elevated predicted risk, even within the same neighborhoods. This finding is consistent with theoretical claims that maintenance outcomes cluster at the portfolio level rather than merely reflecting neighborhood poverty. Cross-program validation also exposes important limitations. In a separate feature-based experiment, an XGBoost model trained on federally inspected Section 8 properties achieved near-random performance when applied to the available Low-Income Housing Tax Credit sample, with an AUC of 0.515. Because this experiment did not directly evaluate the GAT, it does not determine whether graph-based representations would improve cross-program transfer. The result nevertheless cautions against applying models trained on U.S. Department of Housing and Urban Development-inspected properties to other program contexts without program-specific validation. Integrating predictions with economic pressure measures, we identify properties facing compounded threats, including simultaneous physical distress, market exit pressure, contract expiration, and advanced age. These properties concentrate geographically, creating opportunities for targeted intervention. We contribute methodologically by demonstrating what graph neural networks can accomplish within an administratively coherent setting while identifying important limits to cross-program transfer across fragmented housing policy landscapes.