New AI method maps city streets' 'spectral fingerprints'

Piotr C. Kaminski from the University of Warsaw has developed a novel machine-learning approach to quantify urban street networks using spectral graph theory. The method computes 'spectral fingerprints'—fixed-dimensional kernel-density representations of graph-Laplacian eigenvalue distributions for each street's k-hop neighborhood. This addresses a longstanding limitation in space syntax research, where traditional centrality metrics (integration, choice) fail to distinguish streets with identical scores despite radically different morphological contexts.

The innovation lies in two size-adjusted spectral metrics: the Mesh Index (MI), a normalized spectral entropy measuring street fabric complexity, and the Connectivity Resilience Index (CRI), based on algebraic connectivity (Fiedler value). When applied to Poznań's 1,908 streets, the descriptors proved remarkably robust to hyperparameter variations (Spearman ρ ≥ 0.98) while maintaining scale-dependence in neighborhood radius. Crucially, MI showed near-orthogonality to integration (r=0.06), capturing information inaccessible to classical centrality metrics. The method successfully delineated morphological tissue types without supervision and correlated with functional diversity of street-adjacent activities (r~0.19) in OpenStreetMap data.