Domain-Adversarial Transfer Learning for Urban Crime Forecasting: A Case Study of the Surrey—Langley SkyTrain Expansion

This paper investigates domain-adversarial transfer learning for small-area crime forecasting around the Sur-rey-Langley SkyTrain (SLS) expansion in British Columbia, Canada. A two-stage pipeline is proposed: it first trains a Domain-Adversarial Neural Network (DANN) on multi-city data (Surrey and Coquitlam) to learn domain-invariant representations from seasonal-average crime outcomes, coupled with rich covariates: socioeconomic and demographic indicators, transit accessibility, and region-popularity. The pipeline then freezes the extracted features and fits a LightGBM model to predict spatiotemporal crime patterns in the target city (Langley). Timeforward, cross-city evaluations show that DANN-derived features yield better generalization than city-specific models, pooled nonadversarial baselines, and LightGBM trained on raw inputs. Finally, scenario-based forecasts were executed for Langley under simulated post-SLS transit accessibility, producing fine-grained risk maps that highlight potential shifts near the future corridor and station areas. The results indicate that pairing adversarially learned, domain-robust representations with a strong tabular learner offers an effective strategy for transferring knowledge across municipalities and forecasting infrastructure-induced changes in urban safety. A key contribution of this work lies in demonstrating that such a transfer-learning pipeline can support planning decisions for data-scarce municipalities preparing for major transit expansions, where no local historical analogue exists.