Source-linked AI summary
CASTANET: Causality-Aware Spatio-Temporal Adversarial Network Using Traffic Incident Effects
Toshiya Kitahara, Ryu Shirakami, Koh Takeuchi, Hisashi Kashima
TL;DR
Incident-driven congestion is hard to forecast because incidents are sparse, biased in space and time, and have heterogeneous effects across traffic contexts. CASTANET combines spatio-temporal graph neural networks with causal treatment-effect estimation and balanced representations to use incident records while mitigating selection bias. On Tokyo traffic and accident data, it reduces RMSE by 4.0% overall, 10.1% on incident-conditioned evaluation, and up to 14.55% under severe congestion.
Problem
Incident-driven congestion remains difficult to forecast because incidents are extremely sparse, spatially and temporally biased, and heterogeneous in their traffic impacts.
Method
CASTANET integrates an STGNN with treatment-effect estimation, balanced representation learning, and adversarial representation learning for incident-aware congestion prediction.
Results
4.0% lower overall RMSE, 10.1% lower incident-conditioned RMSE, and 14.55% higher improvement under severe congestion are reported on Tokyo data.
Takeaways & Limitations
Modeling incidents as non-random interventions and balancing representations improves queue-length prediction, particularly for post-incident and severe-congestion conditions.
Abstract
from arXiv · showhide
Predicting non-periodic traffic congestion caused by sudden incidents (e.g., accidents and road damage) is crucial for advanced intelligent transportation systems. However, incident-driven congestion is difficult to forecast because incidents are extremely sparse, occur at specific times and locations, and have heterogeneous impacts depending on the traffic context. While recent deep learning approaches have significantly improved periodic traffic forecasting, their performance on non-periodic congestion remains limited, partly because incident records are not explicitly incorporated and their occurrence is strongly biased in space and time. To address these challenges, we propose CASTANET, which integrates spatio-temporal graph neural networks and causal treatment effect estimation to utilize incident records while mitigating selection bias. Experiments on real-world traffic data and accident records from Tokyo, which we treat as incidents, show that CASTANET reduces RMSE by 4.0% overall compared to the best baseline and by 10.1% on incident-conditioned evaluation, with gains reaching 14.55% under severe congestion.
I. INTRODUCTION
Incident-driven congestion is difficult to forecast because impacts vary with traffic and road context, while incident records are sparse and spatially and temporally biased. CASTANET addresses these challenges by combining spatio-temporal graph modeling with causal treatment-effect estimation and reports improvements on Tokyo data.
- Challenges: Incident impacts are heterogeneous: one incident sharply reduces speed and flow while increasing queue length, whereas another produces no clear traffic changes.These contrasting outcomes depend on the incident, traffic conditions, and local road-network structure.
- Challenges: Incidents occur in only 0.01% of samples and are non-uniformly distributed across Tokyo’s times and locations.Their occurrence is also associated with conditions that may independently affect congestion, creating confounding risks.
- Challenges: Naively adding incident indicators can attribute congestion to incidents even when similar congestion would have occurred without them.The paper frames this issue as selection bias in incident-driven prediction.
- Approach: CASTANET combines an STGNN with balanced and adversarial representation learning to model traffic context while reducing dependence on factors that make incidents more likely.The model uses traffic history, incident history, road structure, and current incident presence or absence.
- Results: 4.0% lower overall queue-length RMSE is reported against QTNet, while incident-conditioned RMSE is 10.1% lower than CASTANET:NoARL1.Under severe congestion, CASTANET achieves a 14.55% improvement in incident RMSE against state-of-the-art models.
- Results: Experiments use seven months of Tokyo traffic and incident records and indicate especially large gains in post-incident and severe-congestion scenarios.The evaluation covers approximately two thousand road segments.
II. RELATED WORK
The paper reviews data-driven traffic modeling and causal treatment-effect estimation, then defines the traffic-network inputs and forecasting objective used here.
- Traffic engineering models simulate flow under theoretical assumptions, whereas real-world traffic data often violate those assumptions.
- Treatment effect estimation quantifies outcome changes attributable to treatment while accounting for non-random treatment assignment and missing counterfactuals.
- Balanced representation learning reduces selection bias by making treatment and control representations similarly distributed, improving factual-outcome prediction accuracy.
- The road network is represented by an adjacency matrix W, while traffic and auxiliary features are concatenated into Xt.
- The model uses traffic and incident histories to predict queue length over the next T′ time steps.
- Parameters θ̂ are estimated by minimizing empirical risk with a loss function ℓ.
IV. CASTANET
The paper frames non-uniform incident occurrence across time and space as a problem addressed through treatment effect estimation.
- Incidents are not uniformly distributed across time and space, motivating a treatment effect estimation framework.
A. Causal Viewpoint
CASTANET treats current incidents as non-random treatments and uses balanced spatio-temporal representations to predict future traffic while mitigating incident-selection bias.
- A. Causal Viewpoint: Traffic history Xt, incident history At−1, and road network W are covariates; current incident status at is treatment; future traffic Yt is outcome.
- A. Causal Viewpoint: The causal graph specifies that future traffic depends on traffic history, incident history, road structure, and current incidents, whose occurrence is conditioned on recent covariates.
- A. Causal Viewpoint: The framework assumes consistency, positivity, and sequential strong ignorability, with incidents assumed independent across road segments for simplicity.
- A. Causal Viewpoint: CASTANET integrates STGNNs with treatment effect estimation to incorporate incidents into future-traffic prediction over spatial traffic relationships.
- A. Causal Viewpoint: Incident status is embedded into a learnable vector space, producing historical and current incident embeddings.
- A. Causal Viewpoint: The STGNN encoder extracts Φt from traffic history, incident-history embeddings, and W; Φt predicts future outcomes while being less predictive of current incidents to mitigate bias.
- A. Causal Viewpoint: Graph WaveNet combines traffic and incident-history embeddings, using temporal convolutions and graph convolutions with predefined and learned adjacency structures.
C. Incident-Aware Decoder
The incident-aware decoder combines current incident information with the latent representation to predict future traffic and reconstruct current incident status.
- C. Incident-Aware Decoder: The outcome predictor uses Φt and the current incident embedding h(at), propagating incident information to neighboring road segments.
- C. Incident-Aware Decoder: For queue-length prediction, the decoder predicts speed, flow, and correction terms before applying a queueing-theory layer.
- C. Incident-Aware Decoder: A treatment classification network GA predicts current incident status from Φt using multilayer perceptrons.
D. Adversarial Representation Learning (ARL)
CASTANET uses weighted outcome prediction and counterfactual domain confusion to address sparse, biased incident records while preserving focus on queue-length forecasting.
- The model uses counterfactual domain confusion to make latent representations less predictive of current incident status.This is intended to reduce reliance on factors associated with incident occurrence under specific conditions.
- The outcome predictor targets queue length and uses speed and flow as auxiliary outputs.
- Queue-length losses are weighted to prevent abundant zero-queue samples from dominating training.The weighting also adjusts emphasis between moderate and severe queues through δ1 and δ2.
- A weighted cross-entropy objective trains the treatment classification network while handling incident rarity.
- The treatment-classification outputs are encouraged toward a uniform distribution to increase unpredictability.
- The domain-confusion strength is controlled by the hyperparameter λConf.
V. EXPERIMENTS
The experiments evaluate CASTANET's queue-length prediction against baselines, component ablations, predictive behavior, and scalability on larger road networks.
- The experiments assess whether CASTANET outperforms state-of-the-art baselines, particularly in post-incident scenarios.
- The study tests whether each CASTANET component contributes to prediction performance.
- The experiments investigate how the proposed model predicts traffic congestion.
- The study evaluates how the model scales to a significantly larger road network.
A. Dataset
The dataset combines Tokyo traffic measurements and accident records across a large road network, with chronological splits and incident-aware forecasting inputs.
- The dataset contains seven months of Tokyo traffic and accident data collected from May 1 through November 30, 2022.
- The target network includes 2,032 central-Tokyo road segments spanning 1,468 km.
- Traffic measurements include queue length, speed, and flow for each road segment.
- The traffic control system identifies congestion from sensor-based density estimates exceeding a predefined threshold and uses it to calculate queue length.
- Five datasets use 3- through 7-month periods, each split chronologically into training, validation, and test sets at 7:1:2.
- The input uses the previous 12 timestamps and current incident status to predict queue length for the next 12 timestamps.The historical window and forecast window each represent one hour.
B. Metrics
The evaluation uses RMSE, incident-conditioned RMSE, and queue-length subsets designed to expose performance beyond the abundant zero cases.
- RMSE measures congestion prediction performance.
- Table I reports RMSE and RMSE (Incident) for one-hour-ahead queue-length predictions.Bold and underlined values identify the best and second-best scores, respectively, while parentheses show standard deviation.
- Incident-conditioned RMSE is computed on data points where the incident indicator equals one.
- The test set is divided into All, Non-zero, Top 10%, and Top 5% subsets based on queue length.
- These subsets address the risk that always predicting zero could appear strong on zero-inflated queue-length data.
C. Baseline models
CASTANET is evaluated against conventional and state-of-the-art traffic predictors for one-hour-ahead queue-length forecasting, including incident-conditioned performance and ablations.
- Baseline models: The evaluation compares CASTANET with four conventional models and seven state-of-the-art traffic prediction models.The baselines include HA, Lasso, Ridge, XGBoost, DCRNN, GWNet, AGCRN, MegaCRN, QTNet, PDG2Seq, and STAEFormer.
- Evaluation setup: All models receive concatenated information and predict queue length for the next hour.For models other than QTNet, the loss uses queue-length terms while excluding the weight coefficient r_i,k.
- CASTANET variants: CASTANET is also compared with NoARL1 and NoARL2 variants that use alternative reweighting strategies without adversarial representation learning.NoARL1 applies inverse-propensity weighting, whereas NoARL2 uses inverse empirical incident frequency.
- Training: Hyperparameters are tuned on the validation dataset to improve queue-length accuracy.The implementation includes fixed embedding dimensions, loss weights, and curriculum learning for training the outcome predictor.
- Results: 4.0% lower RMSE is achieved by CASTANET than the best baseline in the All setting, while RMSE (Incident) is reduced by 10.1% versus the best competing method.Under severe congestion, CASTANET improves RMSE (Incident) by 14.55% over state-of-the-art models; ARL outperforms both reweighting variants.
F. Ablation Study
The ablation study tests incident embedding, the QT-layer, and adversarial representation learning, while qualitative predictions compare factual and counterfactual incident inputs. Each component contributes to performance, with ARL most important for incident-conditioned accuracy.
- F. Ablation Study: The ablation compares CASTANET with variants removing incident embedding, the QT-layer, or adversarial representation learning.The variants are CASTANET-IE, CASTANET-QT, and CASTANET-ARL.
- F. Ablation Study: All ablation variants show performance declines relative to CASTANET, indicating contributions from each component.The comparison uses one-hour-ahead RMSE and RMSE (Incident) averaged across datasets (a)–(e).
- F. Ablation Study: ARL contributes most to the improvement, particularly for RMSE (Incident), by mitigating bias and overfitting in the incident distribution.The authors expect this bias mitigation to improve generalization on test data.
- G. Qualitative Evaluation: Factual predictions using incident inputs are more accurate than counterfactual predictions with the current incident indicator set to zero.CASTANET produces predictions every five minutes up to one hour ahead, with incident time marked at 0 minutes.
- G. Qualitative Evaluation: The incident indicator has varying effects across cases, showing that CASTANET adjusts responses according to traffic context rather than applying a fixed rule.The qualitative evaluation compares CASTANET factual and counterfactual outputs, with STAEFormer shown for comparison.
- Computational Evaluation: CASTANET and STAEFormer have comparable inference and training times on the Tokyo dataset.CASTANET requires 0.02 seconds for multi-horizon prediction and 641 minutes for 84 training epochs; STAEFormer requires 0.05 seconds and 656 minutes for 20 epochs.