GeoCTGAN generates privacy-aware geosynthetic event data

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

High-resolution spatial event data are valuable for urban research but difficult to release because of privacy and disclosure risk. GeoCTGAN combines road-network discretization, DeepWalk topology embeddings, globally pooled conditional CTGAN generation and Capacitated k -Nearest Neighbour (CaNN) reconstruction to produce privacy-aware geosynthetic data. Using crime data for the City of Toronto, GeoCTGAN-Full, prioritizing micro-scale spatial fidelity for restricted use, and GeoCTGAN-Priv, incorporating additional privacy protections for more open release, preserved broad spatial structure in the real data, as evidenced by KDE-surface correlations of 0.9976 and 0.9913 and Dissemination Area (DA) count correlations of 0.8919 and 0.8491, respectively. DA maps and LISA classifications were highly concordant with held-out data, while Global Moran’s I retained the same positive spatial-autocorrelation pattern. Temporal evaluation showed close preservation of crime-category and hour-of-day composition across month boundaries. Privacy diagnostics were also strong, with DCR Overfitting Protection score near 0.99 and Disclosure Protection Estimate near 0.73. GeoCTGAN provided a stronger spatial-fidelity and privacy balance than most tabular and spatial-resampling benchmarks. Ablations showed that capacity-constrained CaNN reconstruction was especially important, with topology-based components contributing to finer-scale structure. Multi-seed robustness and unchanged-configuration application to a supplementary second city analysis (City of Ottawa) supported reproducibility and cross-city transferability.