Related Experiment Video
Updated: Jul 16, 2026
14:55
Evaluating the Effect of Roadside Parking on a Dual-Direction Urban Street
Published on: January 20, 2023
Auditing Road-Segment Speed Forecasting Under Sparse Mobile Probe Sensing: A Mask-Consistent Support-Chain Analysis.
Dingxin Wu1, Zheng Xu2, Zhiyuan Wang1
1Faculty of Transportation Engineering, Huai'an University, Huai'An 223001, China.
Sensors (Basel, Switzerland)
|July 15, 2026
Summary
Evaluating urban traffic speed forecasts from sparse GPS data is challenging. This study shows that model performance depends on data availability, not just accuracy metrics, highlighting the need to report evaluation support.
Area of Science:
- Urban mobility analytics
- Traffic flow modeling
- Sensing and data fusion
Background:
- Ride-hailing GPS data offers flexible urban traffic insights but suffers from sparse and uneven coverage.
- This sparsity complicates model evaluation due to misaligned observed targets, predictions, and historical data support.
- Accurate traffic forecasting is crucial for intelligent transportation systems.
Purpose of the Study:
- To audit ultra-short-term road-segment speed forecasting under sparse mobile sensing conditions.
- To introduce and apply a mask-consistent support-chain framework for robust model evaluation.
- To investigate the impact of different evaluation support definitions on forecasting model performance.
Main Methods:
- Utilized a three-day GPS dataset aggregated into 5-minute speed observations for 1970 road segments.
- Implemented an evaluation protocol distinguishing between full test grid, directly observed targets, model-valid prediction support, strict complete-history support, and common-support subsets.
- Compared adaptive graph convolutional recurrent network (AGCRN) and historical mean (HIST_MEAN) baselines under various support conditions.
Main Results:
- The adaptive graph convolutional recurrent network (AGCRN) showed lower mean absolute error (MAE) among full-coverage models.
- The historical mean (HIST_MEAN) baseline achieved lower root mean squared error (RMSE).
- Congestion recall remained below 0.24 for all full-coverage deep models, indicating limited effectiveness in predicting congestion.
Conclusions:
- Model performance is conditional and metric-dependent, with no single model demonstrating universal superiority under sparse sensing.
- Evaluation support must be reported as a primary experimental factor alongside accuracy metrics for sparse mobile probe data.
- Limitations include the short data duration, restricting analysis of temporal and spatial variations.
