TL;DR
Predicting road-level traffic accidents is notoriously hard because accident data is overwhelmingly zero-inflated—most roads experience no crashes on any given day. This paper introduces the STZINB-GNN, a spatial-temporal graph neural network that parameterizes a Zero-Inflated Negative Binomial (ZINB) distribution to jointly forecast micro-level traffic risk and quantify model uncertainty. Evaluated on real-world London data, it significantly outperforms traditional graph networks and Gaussian baselines in pinpointing accident hotspots.
What Problem Does This Paper Address?
Most traffic risk prediction models operate at a coarse, grid-based resolution (e.g., city zones or neighborhoods) and produce simple deterministic point estimates. However, practical road safety and urban planning require fine-grained, road-level predictions.
At the road level, two major challenges arise:
- Extreme Zero-Inflation (Data Sparsity): On a day-to-day basis, the vast majority of individual street segments see zero accidents. Standard regression and deep learning objectives fail because predicting zero everywhere yields deceptively low loss.
- Neglect of Uncertainty: Deterministic black-box models do not communicate confidence. Urban decision-makers need probabilistic estimates to distinguish between predictable low-risk segments and volatile, high-uncertainty zones.
Methods
The authors propose the Spatial-Temporal Zero-Inflated Negative Binomial Graph Neural Network (STZINB-GNN):
- Road Graph Formulation: Roads are represented as nodes $V$ in a graph $G = (V, E)$, connected by an adjacency matrix $A$ reflecting physical network topology.
- Spatial-Temporal Backbone (STGNN):
- A Gated Recurrent Unit (GRU) functions as the temporal encoder, capturing historical risk dynamics over previous time windows.
- A Graph Attention Network (GAT) functions as the spatial encoder, modeling spatial dependencies and spillover effects across neighboring road segments.
- Probabilistic ZINB Output: Instead of outputting a scalar risk score, the STGNN outputs the parameters of a Zero-Inflated Negative Binomial distribution:
- $\pi$: the probability of structural zero (zero-inflation factor).
- $n$ and $p$: the dispersion/success and failure parameters of the Negative Binomial distribution.
- Training Objective: The model is trained end-to-end by minimizing the negative log-likelihood (NLL) of the ZINB distribution over historical risk data.
Datasets
The model was evaluated using an open-source traffic dataset from the Lambeth Borough in London, UK (2019):
- Scale: 5,659 road segments and 1,335 recorded accidents across 2019.
- Risk Formulation: Daily risk incorporates accident frequency, accident severity, and first- and second-order spatial spillover onto neighboring segments.
- Sparsity: The dataset exhibits an extreme zero-inflation rate of 98.7% (only 1.3% of road segments have non-zero risk on any given day).
Key Insights and Ideas
- Separating True Zeros from Fluctuations: Traditional Gaussian and standard Negative Binomial models struggle with extreme sparsity. By introducing the parameter $\pi$, STZINB explicitly separates structural non-risk roads from roads where accidents simply haven't happened yet.
- Entropy-Calibrated Risk Targeting: Instead of just taking the top predicted values, the authors use Kullback-Leibler Divergence (KLD) and information entropy to measure uncertainty. By filtering for roads with high predicted risk and low entropy (high confidence), precision in identifying actual accident sites jumps dramatically.
Key Results
The authors evaluated STZINB against three baselines—Spatial-Temporal GNN with Gaussian Distribution (STG), Spatial-Temporal GNN with standard Negative Binomial (STNB), and Historical Average (HA)—across two time horizons: long-term (14-day history predicting 14 days ahead) and short-term (7-day predicting 7 days):
- Higher Sparsity Accuracy: STZINB achieved a True-Zero Rate (ZR) of 0.641 (long-term) and 0.579 (short-term), compared to 0.199 and 0.102 for the Gaussian baseline (STG).
- Superior Hotspot Detection: On the Hit Rate metric for the top 20% highest-risk roads with low uncertainty (HR20%), STZINB reached 61.8% (long-term) and 57.5% (short-term). This substantially outperformed STNB (44.7% / 44.2%) and STG (26.7% / 30.1%).
- Lower Prediction Error: STZINB consistently achieved lower MAE (0.054) and RMSE (0.119) in long-term forecasting compared to STNB (MAE 0.080, RMSE 0.139) and STG (MAE 0.118, RMSE 0.183).
Limitations and Future Work
- Continuous vs. Count Modeling: The Negative Binomial distribution is theoretically designed for discrete integer counts. However, risk values derived from combining accident severity and neighborhood spillover weights become continuous, non-binary values. The authors acknowledge that a discrete NB distribution may not be mathematically ideal for these continuous risk values.
- Future Directions: Developing and testing alternative continuous or semi-continuous zero-inflated distributions (such as zero-inflated Gamma or Tweedie distributions) that can more naturally handle non-integer, skewed risk indices.
Summary updated