XTraffic: A Dataset Where Traffic Meets Incidents with Explainability and More

Xiaochuan Gou1*, Ziyue Li2* , Tian Lan3, Junpeng Lin3, Zhishuai Li4, Bingyu Zhao5, Chen Zhang3, Di Wang1, Xiangliang Zhang1,6

1King Abdullah University of Science and Technology, Saudi Arabia
2University of Cologne, Germany
3Tsinghua University, China
4Institute of Automation, Chinese Academy of Sciences, China
5Vienna University of Technology, Austria
6University of Notre Dame, USA

DatasetDataset Demo CodeDemo Code

Dataset Description:

Historically, research in traffic and incidents has proceeded along two distinct but intrinsically linked tracks. The traffic domain has focused on enhancing deep learning models to incrementally improve prediction accuracy, while the incident track has predominantly concentrated on isolated studies of incident risks and patterns. For the first time, our XTraffic dataset integrates these two tracks both spatially and temporally across a comprehensive regional scale, encompassing 16,972 traffic nodes for the entire year of 2023. The dataset includes detailed time-series data on traffic flow, lane occupancy, and average vehicle speed, as well as meticulously aligned records of incidents across seven different classes, synchronized with the traffic data. Each node also features extensive physical and policy-level meta-attributes of lanes.

Descriptive Analysis

DescriptiveAnalysis

Experiments

Traffic Forecasting

Traffic forecasting is the task of predicting future traffic conditions, such as traffic volumes, speeds, and occupancy rates, at various locations within a transportation network using historical traffic data and other relevant factors.

Test Model 5 Mins (t=1) 15 Mins (t=3) 30 Mins (t=6)
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE
General LSTM 12.58 11.81 21.45 15.41 14.21 26.29 18.68 18.03 31.54
ASTGCN 12.45 13.11 20.90 14.59 13.66 23.10 16.03 15.56 27.82
DCRNN 11.90 11.82 20.47 13.41 12.92 23.79 14.84 14.32 26.74
AGCRN 12.54 12.56 22.65 13.55 13.18 25.27 14.62 14.24 27.92
GWNET 11.99 11.85 20.30 13.53 12.87 23.44 14.88 14.15 26.05
STGODE 12.75 13.26 21.66 14.12 14.57 24.64 15.50 16.34 27.43
DSTAGNN 13.18 12.15 21.93 16.37 18.82 27.41 19.99 19.97 33.73
D2STGNN 12.18 12.00 21.30 13.48 13.20 24.30 14.90 14.27 27.28
Incident LSTM 14.17 10.13 23.75 17.41 15.38 29.43 20.93 14.33 34.05
ASTGCN 14.06 10.55 23.22 16.42 15.06 27.63 18.40 12.59 30.48
DCRNN 13.62 9.86 23.04 15.36 14.35 26.73 16.92 11.69 29.38
AGCRN 14.48 10.98 25.41 15.96 14.78 28.78 17.21 11.78 31.42
GWNET 13.73 10.44 22.90 15.60 14.50 26.73 17.15 11.49 29.07
STGODE 14.50 10.71 24.49 16.20 15.19 27.69 17.55 12.29 30.17
DSTAGNN 14.79 10.57 24.22 18.24 17.91 30.48 21.95 15.38 35.67
D2STGNN 13.73 10.05 23.30 15.51 14.22 27.29 17.03 11.46 30.23

Incident Analysis

Incident classification is the task of identifying and categorizing incidents, such as accidents or hazards, based on the analysis of traffic time series data from a specific road segment.

Methods Speed channel-only Occupancy channel-only Flow channel-only All channels mixed
Acc Precision Recall Acc Precision Recall Acc Precision Recall Acc Precision Recall
DT 41.6% 41.5% 41.5% 40.4% 40.2% 40.2% 39.4% 39.3% 39.3% 41.6% 41.4% 41.5%
TS2Vec 36.6% 36.2% 36.2% 36.6% 36.5% 36.4% 37.3% 37.0% 37.0% 37.3% 37.0% 37.0%
gMLP 41.3% 41.2% 41.1% 38.4% 38.3% 38.3% 37.3% 37.2% 37.2% 41.6% 41.5% 41.5%
Sequencer 35.8% 35.8% 35.6% 35.6% 35.3% 35.2% 34.1% 33.9% 33.9% 40.3% 40.2% 40.2%
OmniScaleCNN 35.7% 35.1% 35.1% 36.9% 36.3% 36.3% 37.0% 36.8% 36.8% 40.9% 40.8% 40.8%
PatchTST 38.3% 38.1% 38.1% 39.0% 38.6% 38.7% 39.5% 39.3% 39.3% 39.4% 39.4% 39.3%
FormerTime 35.9% 31.0% 33.4% 41.0% 41.1% 40.8% 37.8% 38.2% 37.3% 40.5% 40.5% 40.1%

Global Causal Analysis with Meta Data

We learned the DAG among meta-features, incidents, and traffic indexes. Also, we give a factual explanation for selected edges.

Global Causal Analysis Figure 1

Local Causal Analysis

We conduct local causal analysis on a real case from XTraffic

Local Causal Analysis Figure 1

Additional Information

Dataset License: Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

Code License: MIT

Paper Citation

Upcoming soon...