A Spatio-Temporal Approach with Self-Corrective Causal Inference for Flight Delay Prediction

TL;DR

Flight delays propagate through complex inter-airport interactions, challenging accurate forecasting. The authors develop CausalNet, a self-corrective spatio-temporal graph neural network that builds Granger-based causal graphs among airports, refines them with a trainable correction module, and fuses causal and geographic information to capture heterogeneous spatial effects. Temporal dynamics are modeled with Long-Gate Recurrent Units in an encoder-decoder framework, enabling robust multi-step predictions. Experiments on 74 busiest Chinese airports show clear gains over strong baselines, with ablations confirming the value of self-corrective causality and heterogeneity-aware fusion for practical air-traffic management insights.

Abstract

Accurate flight delay prediction is crucial for the secure and effective operation of the air traffic system. Recent advances in modeling inter-airport relationships present a promising approach for investigating flight delay prediction from the multi-airport scenario. However, the previous prediction works only accounted for the simplistic relationships such as traffic flow or geographical distance, overlooking the intricate interactions among airports and thus proving inadequate. In this paper, we leverage causal inference to precisely model inter-airport relationships and propose a self-corrective spatio-temporal graph neural network (named CausalNet) for flight delay prediction. Specifically, Granger causality inference coupled with a self-correction module is designed to construct causality graphs among airports and dynamically modify them based on the current airport's delays. Additionally, the features of the causality graphs are adaptively extracted and utilized to address the heterogeneity of airports. Extensive experiments are conducted on the real data of top-74 busiest airports in China. The results show that CausalNet is superior to baselines. Ablation studies emphasize the power of the proposed self-correction causality graph and the graph feature extraction module. All of these prove the effectiveness of the proposed methodology.

Figures (6)

• Figure 1: The architecture of the proposed self-corrective spatio-temporal graph neural network.
• Figure 2: Self-Causal Correction Module. This figure shows the specific process of the Self-Causal Correction Module. First, the causal graph is aggregated through GCN to aggregate node features. The output $HC^t$ will perform matrix operations with two transformation matrices $E_1$ and $E_2$ to obtain comparison matrices $\rho_1$ and $\rho_2$. By calculating the similarity between the comparison matrices, the degree of correction needed for the causal graph can be determined, and a correction mask $CM^{t}$ is constructed accordingly. The blue edges in the correction mask represent the original causal relationships, and the red edges represent the causal relationships that need to be corrected.
• Figure 3: Compared with the improvement rates of different algorithms in MAE and RMSE. (a)(b)(c) shows the proportion of different algorithms' MAE exceeding our model within 1 to 3 hours, and (d)(e)(f) shows the proportion of different algorithms' RMSE exceeding our model within 1 to 3 hours.
• Figure 4: Performance of the ablation experimental model on MAE and RMSE. (a)(b)(c) shows the proportion of MAE of different CausalNet removing some components that is higher than the original model within 1 to 3 hours. (d)(e)(f) shows the proportion above RMSE from 1 to 3 hours.
• Figure 5: Causal correction result. This figure shows the differences between the original causal graph and the corrected causal graph from four scales: year, month, week, and day.