GT-CausIn: a novel causal-based insight for traffic prediction

TL;DR

This work tackles traffic speed forecasting by leveraging causal structure discovery to inform spatial-temporal modeling. It defines legitimate traffic causal variables and uses Iterative Causal Discovery (ICD) to learn a $30\times 30$ adjacency that captures interactions among a node, its first- and second-order neighbors, and speed variation signals, then integrates this causal insight with directed graph diffusion and a temporal convolutional network (TCN) in the GT-CausIn architecture. The approach yields state-of-the-art performance on mid-term and long-term horizons on PEMS-BAY and METR-LA, with ablations confirming the value of neighbor-aware causal inputs and the GT-block depth. The work contributes a practical framework for incorporating causal knowledge into spatiotemporal traffic forecasting, offering improved accuracy and interpretable insights for real-world smart-city applications.

Abstract

Traffic forecasting is an important application of spatiotemporal series prediction. Among different methods, graph neural networks have achieved so far the most promising results, learning relations between graph nodes then becomes a crucial task. However, improvement space is very limited when these relations are learned in a node-to-node manner. The challenge stems from (1) obscure temporal dependencies between different stations, (2) difficulties in defining variables beyond the node level, and (3) no ready-made method to validate the learned relations. To confront these challenges, we define legitimate traffic causal variables to discover the causal relation inside the traffic network, which is carefully checked with statistic tools and case analysis. We then present a novel model named Graph Spatial-Temporal Network Based on Causal Insight (GT-CausIn), where prior learned causal information is integrated with graph diffusion layers and temporal convolutional network (TCN) layers. Experiments are carried out on two real-world traffic datasets: PEMS-BAY and METR-LA, which show that GT-CausIn significantly outperforms the state-of-the-art models on mid-term and long-term prediction.

Figures (11)

• Figure 1: Example of deriving speed variation of each node with 5 timestamps as input.
• Figure 2: System architecture for the Graph Spatial-Temporal Network Based on Causal Insight. Time series are fed into the causal insight layer and a series of GT blocks. Skip connections are used at the end of each GT block to guard useful information. Periodic features are finally merged to make predictions.
• Figure 3: A brief illustration of the causal insight layer. Node features are first concatenated with neighbor embedding features, then pass to an attention layer to reveal interaction influence. Three matrices are followed to learn individual influence on each node.
• Figure 4: Periodic speed changes of stations of PEMS-BAY within one week.
• Figure 5: Parameter analysis of $L$ (number of stacking GT layers) on METR-LA dataset