Abstract: Due to the complex and dynamic traffic contexts, the interpretability and uncertainty of traffic forecasting have gained increasing attention. Significance testing is a powerful tool in statistics used to determine whether a hypothesis is valid, facilitating the identification of pivotal features that predominantly contribute to the true relationship. However, existing works mainly regard traffic forecasting as a deterministic problem, making it challenging to perform effective significance testing. To fill this gap, we propose to conduct Full Bayesian Significance Testing for Neural Networks in Traffic Forecasting, namely ST-nFBST. A Bayesian neural network is utilized to capture the complicated traffic relationships through an optimization function resolved in the context of aleatoric uncertainty and epistemic uncertainty. Thereupon, ST-nFBST can achieve the significance testing by means of a delicate grad-based evidence value, further capturing the inherent traffic schema for better spatiotemporal modeling. Extensive experiments are conducted on METR-LA and PEMS-BAY to verify the advantages of our method in terms of uncertainty analysis and significance testing, helping the interpretability and promotion of traffic forecasting.
Introduction
Traffic forecasting is a critical component of Intelligent Transportation Systems (ITS), yet traditional deep learning models often operate as "black boxes," providing deterministic predictions without quantifying uncertainty or explaining the underlying logic. This lack of interpretability is problematic in dynamic traffic environments where sensor noise or unexpected incidents can lead to high-stakes errors. While significance testing is a standard statistical tool for identifying pivotal features, its application to complex, non-linear neural networks has been mathematically challenging.
The ST-nFBST Framework
To address these gaps, the authors propose ST-nFBST (Full Bayesian Significance Testing for Neural Networks in Traffic Forecasting). This framework transforms traffic forecasting into an interpretable statistical process through three core components.
First, the model employs Bayesian Spatial-Temporal Modeling, using Bayesian Neural Networks (BNNs) where weights are represented as probability distributions rather than fixed values. It leverages Diffusion Convolutional Gated Recurrent Units (DCGRU) to capture the spatial "flow" of traffic across a graph. Second, a dual-decoder architecture separates aleatoric uncertainty (inherent data noise) from epistemic uncertainty (model parameter uncertainty). Finally, the Grad-nFBST module introduces a gradient-based significance test. By analyzing the gradient of the output with respect to the input, the model calculates "Bayesian evidence" values to statistically determine which historical time steps or sensors significantly drive a prediction.
Experimental Results and Insights
The model was evaluated on the METR-LA and PEMS-BAY benchmarks. ST-nFBST consistently outperformed state-of-the-art baselines, including DCRNN and DeepSTUQ. For instance, on METR-LA at a 1-hour horizon, it achieved an MAE of 3.52, surpassing the previous best Bayesian model.
Beyond accuracy, the framework provides deep insights into traffic dynamics. Significance testing revealed that short-term forecasts (15-min) rely heavily on the most recent 15–30 minutes of local data. As the horizon extends to 1 hour, the model's reliance on its own history decreases, and it begins to incorporate broader network states, reflecting spatial propagation.
Uncertainty quantification also serves as a powerful monitoring tool. As shown in the experimental analysis, spikes in epistemic uncertainty often coincide with sudden changes in road conditions, such as accidents. Furthermore, the results indicate that aleatoric uncertainty (sensor noise) typically dominates epistemic uncertainty in traffic contexts. By pruning "insignificant" features identified through testing, the authors demonstrated that retraining on "cleaned" signals further improves model stability and accuracy.
Conclusion
ST-nFBST offers a robust path toward interpretable and reliable traffic forecasting. While the computational cost of Bayesian inference remains a challenge for massive networks, the ability to "interrogate" a model's decisions provides invaluable transparency for ITS applications. Future work may extend this framework to other spatial-temporal domains like air quality or power grid management.
Summary updated