Trustworthy Spatio-Temporal Graph Learning for Urban Event Forecasting: A Systematic Survey of Explainability, Fairness, and Their Intersection

Spatio-temporal graph neural networks (STGNNs) are widely used for urban event forecasting, including traffic prediction, crime hotspot estimation, air-quality monitoring, and epidemic surveillance. However, their adoption in public governance requires both explainability and fairness, in addition to predictive accuracy. Although existing surveys have separately examined STGNN architectures, graph explainability, and graph fairness, their intersection remains largely unexplored. This gap is significant because relational structures can propagate demographic and structural biases, while access to reliable explanations may vary across sensitive groups. Following a systematic review of publications from 2016 to 2026, this study organizes STGNN architectures for urban forecasting, develops a taxonomy of intrinsic and post-hoc explainability methods, and categorizes fairness notions, bias sources, and debiasing strategies for dynamic urban graphs. It also examines joint explainability–fairness approaches, including temporal fairness stability and explanation disparity, and consolidates relevant datasets and evaluation protocols. The findings show that fewer than one in ten fairness-aware or explainable STGNN studies addresses both properties, while none before 2024 optimized them jointly. We conclude with a research agenda addressing temporal explanation benchmarks, dynamic fairness, causal grounding, scalability, and data-sparse Global-South cities.