CIPHER: Interpretable Price Inference by Fusing Mobility Networks and ZIP-Level Socioeconomics

Urban housing prices arise out of the nuanced interplay of neighbourhood socioeconomic, land-use, and mobility. This paper presents the Construction–Integration–Prediction–Hypothesis–Explanation–Reproducibility (CIPHER) framework—an interpretable cross-city model combining different types of mobility graphs (parking, metro, and bike) and ZIP-level features within ZIP Code Tabulation Areas (ZCTAs) including parks, income, education, and schools, through a Heterogeneous Graph Transformer (HGT). Here, CIPHER denotes: Construction of a type-aware urban graph; Integration of ZIP context at the node level for early fusion; Prediction with a type-specific multi-task head centered on median price; Hypothesis-driven probes separating structure from parameters via hard edge deletions and value-only scaling; Explanation via Input×Gradient (IXG) and Integrated Gradients (IG) attributions aggregated from nodes to ZCTAs; and Reproducibility through schemas and manifests across cities. Comparisons of San Francisco (SF) and New York City (NY) show CIPHER achieving price prediction accuracy with between 0.77 and 0.81, with stronger performance for metro in NY and for parking in SF. The ancillary population task has poor fit. Discussion of edge removal points out neighbourhood structure dependence, and value-only scaling captures parameter robustness. Attributions at the ZIP-level provide policy-relevant maps distinguishing between corridor-driven and community-driven areas.