Research map / Time series / Causal and generative time series
TraffiDent: A Dataset for Understanding the Interplay Between Traffic Dynamics and Incidents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TraffiDent, a large-scale PeMS-based dataset aligning lane-level traffic indexes (flow, occupancy, speed) for 16,972 sensors with 1.44M incident records and road meta-features, and demonstrates four tasks on it.
Problem
Traffic forecasting and incident analysis are studied separately; existing traffic datasets lack spatiotemporally aligned incident records and detailed road meta-features, and incident datasets lack traffic time series, limiting research on traffic-incident interactions and causality.
Approach
A spatiotemporally aligned traffic-incident dataset from Caltrans PeMS covering 2022-2024 with seven incident classes, physical and policy meta-attributes per lane, and benchmark tasks (post-incident forecasting, incident classification, global and local causal analysis).
Key results
- In post-incident forecasting on Monterey D5, all eight forecasting baselines have higher errors on incident samples than general samples (e.g., DCRNN 5-min MAE 11.90 general vs 13.62 incident).
- Best incident classification accuracy is 41.6% (DT on speed channel and gMLP on mixed channels), versus ~33% random guess for three classes.
- Global causal analysis on four districts with 17 variables uses MM-DAG; local causal analysis uses PCMCI+ on four flow nodes near a Novato interchange incident.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate forecasting on incident-affected samples instead of the general test set | Traffic forecasting baselines · Monterey traffic forecasting, 5–30 min horizons | forecasting error | Hurt |
| Combine speed, occupancy, and flow channels rather than using one channel | Incident classification models · Incident classification | accuracy | Mixed |
| Use speed rather than flow as the input channel | Sequencer · Incident classification | accuracy | Helped |
| Use flow rather than speed as the input channel | OmniScaleCNN · Incident classification | accuracy | Helped |
Surprising findings
- Incident classification from traffic indexes is hard: accuracy peaks near 41%, and short-lived incidents such as car fires and non-injury accidents barely change traffic patterns.
- Fringe-node incidents rise slightly on weekends while hub-node incidents are higher on weekdays, and Friday shows the highest incident counts.
Limitations
- Error bars and statistical significance are not reported for forecasting and classification experiments.
- Forecasting and classification experiments use only subsets (San Bernardino/Monterey D5 for forecasting; nearest-sensor windows for classification), and causal analyses are qualitative without ground truth.
- Incident-to-sensor matching is approximate (freeway name and absolute postmile, with a distance threshold), and missing values are zero-filled in experiments.
Experimental setup
Datasets: TraffiDent PeMS traffic data PeMS incident records
Benchmarks: TraffiDent
Metrics: MAE MAPE RMSE accuracy precision recall
Baselines: LSTM ASTGCN DCRNN AGCRN GWNET STGODE DSTAGNN D2STGNN TS2Vec gMLP Sequencer OmniScaleCNN
Builds on: MM-DAG PCMCI+ PEMS LargeST
Closest papers by meaning
- FT-AED: Benchmark Dataset for Early Freeway Traffic Anomalous Event Detection (NeurIPS 2024 · similarity 0.86)
- RiskOracle: A Minute-Level Citywide Traffic Accident Forecasting Framework (AAAI 2020 · similarity 0.86)
- Graph Neural Networks for Road Safety Modeling: Datasets and Evaluations for Accident Analysis (NeurIPS 2023 · similarity 0.86)
- Towards Enhancing Road Safety in South Carolina Using Insights from Traffic and Driver-Education Data (Student Abstract) (AAAI 2025 · similarity 0.85)
- Accident Anticipation via Temporal Occurrence Prediction (NeurIPS 2025 · similarity 0.85)
- Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.