Tourism-context-aware one-class anomaly detection in urban visitor movement: tourist-held-out validation and a governance framework

Urban destinations increasingly monitor visitor movement, yet it remains unclear whether complex anomaly-detection models add value on small tourism datasets, or how statistical anomalies should be interpreted for destination governance. Five one-class models – Isolation Forest, One-Class Support Vector Machine, K-Means, a standard autoencoder, and a deep autoencoder – were evaluated against a training-free zone rule on 1,000 synthetic records from 284 tourist identifiers, using five-fold GroupKFold that held out complete tourist identifiers and an inner tourist-disjoint validation set for thresholds. Robustness was assessed through a tourist-clustered paired randomisation test, cyclic-hour ablation, threshold analysis, prevalence bootstrapping and standardised sensor-noise tests. Anomaly labels proved largely determined by zone designation: flagging every Restricted- or Sensitive-zone record recovered 107 of 109 anomalies (recall 0.981 ± 0.026; F1 0.545 ± 0.047), and no learned model surpassed this rule on recall or F1. Isolation Forest led on accuracy (0.890 ± 0.027) and ROC-AUC (0.920 ± 0.025). The findings show how an apparently demanding benchmark can chiefly reward recovery of a synthetic labelling rule, and support a relational account of anomaly significance that separates statistical rarity from managerial relevance. Destination managers should therefore treat anomaly scores as prompts for human verification rather than evidence of threat.