Source-linked AI summary
RiskOracle: A Minute-level Citywide Traffic Accident Forecasting Framework
Zhengyang Zhou, Yang Wang, Xike Xie, Lianliang Chen, Hengchang Liu
TL;DR
Minute-level traffic accident forecasting is important for timely urban safety decisions but is challenged by rare accidents, zero-inflated labels, and rapidly changing road-network conditions. RiskOracle combines preprocessing, Multi-task DTGN, and region selection to model dynamic correlations and highlight likely accident subregions. Across two real-world datasets, it outperforms baselines on 30-minute and 10-minute tasks, with a reported 22.49% Acc@20 improvement on NYC.
Problem
Existing accident forecasting is often conducted at hour-level granularity, while minute-level prediction faces dynamic road conditions, rare accidents, and zero-inflated labels.
Method
RiskOracle uses preprocessing strategies, Multi-task DTGN, and multi-scale region selection to model dynamic subregion correlations and identify likely accident areas.
Results
RiskOracle surpasses state-of-the-art solutions on 30-minute and 10-minute tasks and improves NYC Acc@20 accuracy by 22.49% over the best baseline.
Takeaways & Limitations
The framework provides more timely citywide accident-risk forecasts while remaining effective for sparse, sporadic spatiotemporal data.
Takeaways & Limitations
Accident risk types are represented using fixed weights of 1, 2, and 3 for minor, injured, and fatal accidents.
Abstract
from arXiv · showhide
Real-time traffic accident forecasting is increasingly important for public safety and urban management (e.g., real-time safe route planning and emergency response deployment). Previous works on accident forecasting are often performed on hour levels, utilizing existed neural networks with static region-wise correlations taken into account. However, it is still challenging when the granularity of forecasting step improves as the highly dynamic nature of road network and inherent rareness of accident records in one training sample, which leads to biased results and zero-inflated issue. In this work, we propose a novel framework RiskOracle, to improve the prediction granularity to minute levels. Specifically, we first transform the zero-risk values in labels to fit the training network. Then, we propose the Differential Time-varying Graph neural network (DTGN) to capture the immediate changes of traffic status and dynamic inter-subregion correlations. Furthermore, we adopt multi-task and region selection schemes to highlight citywide most-likely accident subregions, bridging the gap between biased risk values and sporadic accident distribution. Extensive experiments on two real-world datasets demonstrate the effectiveness and scalability of our RiskOracle framework.
Introduction
Minute-level accident forecasting matters for timely urban safety decisions, but finer temporal granularity exposes zero-inflated labels and dynamic spatiotemporal challenges that existing approaches overlook.
- Minute-level forecasting supports more timely safe-route recommendations and emergency responses than coarser forecasting periods.
- Existing accident-forecasting studies commonly operate at day or hour levels, limiting their usefulness for emergency conditions.
- Finer spatiotemporal resolution produces zero-inflated training targets, allowing models to predict zeros while rare nonzero accidents have little influence.
- RiskOracle addresses minute-level citywide prediction through preprocessing, Multi-task DTGN modeling, and learned selection of likely accident subregions.
Preliminaries and Problem Defintion
The paper represents a city as an urban graph whose subregions carry static road-network and dynamic traffic features, supporting citywide accident-risk prediction from historical observations.
- The study area is partitioned into m square subregions modeled as vertices of an urban graph, with edges indicating subregion connectedness.
- Affinity matrices retain ρ-connectedness, so nonzero affinities represent subregions with strong traffic correlations.
- Dynamic traffic features combine human-activity intensity, average traffic speed, and accident risk for each subregion and interval.
- The prediction problem uses static road-network features and historical dynamic traffic features to forecast the citywide traffic-accident distribution.
Forecasting
RiskOracle is presented as a three-stage framework spanning data preprocessing, model training, and prediction, with accident risk represented using weighted accident types.
- RiskOracle consists of data preprocessing, model training, and prediction stages.
- The forecasting setup defines minor, injured, and fatal accidents with weights 1, 2, and 3, respectively.
Framework Overview
The framework overview organizes RiskOracle into three sequential stages for preprocessing data, training the model, and producing predictions.
- RiskOracle’s framework overview contains data preprocessing, model training, and prediction stages.
- The stages are presented as the framework’s overall processing flow.
Data Preprocessing
RiskOracle organizes citywide accident prediction around hierarchical regions and transforms sparse, zero-heavy labels for training. It also uses ST-DFM to infer missing traffic information from static, dynamic, and temporal feature interactions.
- Spatial Organization: Hierarchical regions address spatial imbalance by representing fine- and coarse-grained accident distributions across urban and rural areas.The study area is divided into medium-sized rectangular regions containing smaller square subregions.
- Overcoming Zero-inflated Issue: Only 6 accidents occur across NYC during one 10-minute interval, illustrating why minute-level labels are inherently sparse.The paper uses this example to motivate its priori knowledge-based data enhancement strategy.
- Overcoming Zero-inflated Issue: The label transformation makes zero-item regions negative while preserving the ranking of actual accident risks.The logarithmic transformation is designed to make labels discriminating and suitable for network training.
- Complementing Sparse Sensing Data: ST-DFM combines static spatial features, dynamic traffic features, and timestamps to learn feature interactions and high-level representations.Its CIN and DNN modules model vector-wise interactions and higher-level feature combinations.
Multi-task DTGN for Accident Risk Prediction
Multi-task DTGN models minute-level accident risk using dynamic inter-subregion affinities and differential traffic changes. Its auxiliary tasks jointly represent risk, traffic volume, and accident counts.
- Time-varying Overall Affinity: Dynamic overall affinity captures time-varying traffic correlations, including links between distant subregions with accident-related traffic characteristics.The affinity combines dynamic traffic conditions with static spatial and accident-related features.
- Differential GCN: Differential GCN extracts abnormal traffic changes between adjacent intervals to model spatiotemporal propagations and interactions.The differential feature generator complements dynamic traffic features before graph convolution.
- Temporal Perspectives: DTGN uses hourly closeness, daily periodicity, and distant trend to select temporal feature tuples for forecasting.For distant trend, the model samples previous days at 10-day intervals; κ is set to 3.
- Graph Convolution: The graph convolution recursively transforms features with learned kernels, while mean affinity matrices, batch normalization, and Leaky ReLU support stable processing.External timestamps and meteorological data are embedded and fused with the learned representation.
- Multi-task Learning: The multi-task scheme predicts accident-risk distributions alongside traffic volume and coarse-grained accident counts.The auxiliary outputs are fused with risk features to support hierarchical accident-distribution learning.
Hierarchical Most-likely Accident Region Selection
HARS selects likely accident subregions hierarchically rather than applying one citywide risk threshold. It uses learned regional accident counts to adapt the number of selected subregions.
- Motivation: A unified risk threshold is unsuitable because accident coverage and traffic volumes differ between urban and rural areas.The resulting spatial heterogeneity can cause high-risk values to favor downtown regions and overlook relatively risky rural regions.
- Hierarchical Selection: HARS selects k_i highest-risk subregions within each rectangular region, with k_i learned by the auxiliary accident-count task.The resulting set contains the most-likely accident regions while adapting to time and weather changes.
Empirical Studies
RiskOracle is evaluated on two real-world datasets using 30-minute and 10-minute forecasting intervals, regression and spatial-classification metrics, and comparisons with five baselines. It achieves the strongest reported performance across nearly all metrics, remains more robust as temporal granularity becomes finer, and supports adaptive accident-region selection.
- Empirical Studies: Experiments use NYC and SIP data, 30-minute and 10-minute intervals, and MSE, Acc@M, and Acc@K to assess regression and spatial classification.NYC uses taxi-trip volumes as a human-mobility indicator, while SIP provides traffic flows and speeds alongside accident records.
- Comparison Performances: RiskOracle's performance decreases slightly with finer temporal granularity, while baselines decrease sharply because of the zero-inflated issue.The same pattern appears across both datasets, supporting the framework's reported robustness for short-term accident prediction.
- Evaluations on Acc@K and Ablation Studies: RiskOracle outperforms other methods on Acc@K, whose learned K adapts the selected accident-region count to each interval.The authors attribute this flexibility to multi-scale accident-distribution forecasting and hierarchical accident selection.
- Evaluations on Acc@K and Ablation Studies: The integrated RiskOracle consistently outperforms ablated variants at both temporal levels, with time-varying overall affinity and PKDE contributing the largest gains.The ablation removes PKDE, ST-DFM, overall affinity, differential feature generation, and multi-task HARS in turn.
Hyper-parameter Studies
The study examines model settings and visualizes selected 30-minute predictions in NYC. The case study reports spatial agreement with observed accidents and adaptive changes in predicted risk across times and external conditions.
- Hyper-parameter Studies: 53.82% Acc@K is reached with λ1 = 0.8 and λ2 = 1 in the NYC 30-minute parameter study.The study also reports best settings of 9 layers with 384 filters, γ = 0.5, and q = 18 for the examined configurations.
- Case Study: RiskOracle's citywide risk maps show spatial similarity between highlighted predicted subregions and ground-truth accidents at selected 30-minute intervals.The case study visualizes predictions for one day in NYC.
- Case Study: Predicted accidents are rare at 7:00 a.m., increase in the afternoon, and become worse in the evening during the illustrated Sunday sequence.The passage attributes the evening increase mainly to heavy rain and accident-prone road conditions.
Conclusion
RiskOracle addresses minute-level citywide accident forecasting by combining Multi-task DTGN with strategies for sparse sensing, sporadic data, and spatial heterogeneity.
- Conclusion: RiskOracle combines Multi-task DTGN with preprocessing and multi-task strategies for minute-level citywide accident forecasting.The framework targets sporadic spatiotemporal data, insufficient sensing, and spatially heterogeneous accident distributions.
- Conclusion: Differential feature generation and time-varying overall affinity model sporadic data and short-term subregion-wise correlations.
- Conclusion: Experiments on two real-world datasets show that RiskOracle outperforms state-of-the-art solutions.