Traffic flow prediction is a fundamental task in intelligent transportation systems, challenged by dynamic spatial dependencies and multi-scale temporal evolution. Existing graph neural network- and Transformer-based methods directly model node relations in the raw feature space, making it difficult to separate genuine causal dependencies from spurious correlations caused by external factors, which leads to prediction bias and reduced model stability. To address these issues, this study proposes the Interventional Geometric Wavelet Transformer (IGWave), which integrates causal intervention with geometric modeling for spatiotemporal prediction. Specifically, the model applies the stationary wavelet transform (SWT) to decompose traffic sequences and to extract long-term trend features and short-term event features. Furthermore, an Interventional Geometric Attention (IGA) module is incorporated into the attention layer. Within IGA, causal intervention suppresses confounding effects, and a geometric wedge kernel jointly encodes directional and magnitude dependencies to enforce causal consistency and directional robustness during attention computation. A multi-layer IGA encoder then integrates the extracted spatiotemporal features and directly produces multi-step prediction outputs. Experiments on four real-world datasets demonstrate that IGWave surpasses mainstream baselines in accuracy, stability, and cross-domain generalization, establishing a new paradigm for interpretable and geometrically robust causal modeling in complex spatiotemporal systems.
Linlong Chen (Sun,) studied this question.