Authors: Husna Sultana, Irfan Ahmed, Shivani
Abstract: Interpreting the Urban Black Box, the proliferation of sensors and Internet of Things (IoT) infrastructure in Smart Cities has enabled the development of highly accurate Spatio-Temporal Data Mining models, often relying on deep learning architectures like Graph Neural Networks (GNNs), for tasks such as traffic prediction, crime forecasting, and resource management. Despite their high predictive performance, these models remain "black boxes," hindering their adoption by urban planners and emergency services who require transparency and justification for critical operational decisions. This lack of interpretability poses significant challenges to accountability, auditability, and public trust. This paper addresses the critical need for Explainable AI (XAI) in the urban domain by proposing a novel Spatio-Temporal XAI (ST-XAI) Framework designed for Causal Feature Attribution. Our framework leverages a modified version of SHapley Additive exPlanations (SHAP) combined with the inherent spatial and temporal structure of the data to provide granular, instance-based explanations. The proposed methodology focuses on Temporal Attribution: Quantifying the specific influence of various look-back time windows (e.g., data from the last hour vs. data from 24 hours ago) on the current prediction. Spatial Attribution: Identifying and weighting the contributing influence of specific geographic nodes, links, or neighboring zones within the network structure. Causal Inference: Moving beyond mere correlation by prioritizing features that exhibit a strong, temporally preceding impact, providing a more actionable justification for the prediction. We demonstrate the ST-XAI Framework on a smart traffic prediction model, showing how it successfully translates opaque deep learning outputs into clear, human-understandable narratives. The results illustrate that our framework not only validates model efficacy but also acts as a vital debugging tool for city engineers, transforming black-box predictions into accountable and actionable urban intelligence.
DOI: http://doi.org/10.5281/zenodo.17720787
