Integrating Causality with Spatio-Temporal Attention for Accurate Airline Delay Prediction
Abstract
Flight delays can cause serious problems for airlines, passengers, and the economy in general. Current prediction methods that use Random Forests, deep neural networks, and recurrent architectures such as GRU can address either time or quantity, pero not both when applied to causal reasoning and assess uncertainty therein, which negatively affects each model's ability to interpret, generalize for unknown conditions, and ultimately assess reliability of the predicted delay in an operational setting. Causal-Aware Spatio-Temporal Attention Network (CASTAN) is designed as a combined approach to address these challenges of spatio-temporal and causal modeling all in one. Analysts use GraphSAGE-based spatial encoding to encode and capture inter-airport dependencies, with a self-attention temporal encoder to learn long-range sequential patterns of historical delays in addition to traffic and weather factors. A cross-attention fusion mechanism accounts for the dynamic and spatio-temporal contributions to delay. A final causal counterfactual module adds interpretable independence results—helping analysts to assess the contributing factors to delay. Finally, the incorporation of dropout is done in a Bayesian approach to assess uncertainty for each prediction made and generate uncertainty-aware predictions so analysts may assess reliability through levels of confidence or any other metric decided. Results from evaluation of a large-scale U.S. flight dataset compared to traditional baselines demonstrate the predictive power of the model, achieving 96.4% accuracy, RMSE of 4.2, and MAE of 2.9. The CASTAN process has positioned its place as an interpretable, reliable, and operationally informative modeling approach to proactive management of airline delay.