Uncertainty-aware Traffic Prediction under Missing Data

TL;DR

The paper tackles traffic forecasting when some locations lack historical data and argues for explicit uncertainty quantification to support risk-sensitive decisions. It introduces the uncertainty-aware inductive graph neural network (UIGNN), which combines inductive graph learning with a diffusion GCN backbone and an evidential deep learning layer to predict traffic states at both observable and missing locations and to quantify uncertainty per location. Key contributions include a sub-graph sampling training scheme enabling generalization to unseen locations, a diffusion-based prediction backbone, and an uncertainty-aware output that correlates with data availability and supports downstream tasks like active sensing and sensor deployment. Across three real-world datasets, UIGNN outperforms two-step imputations and provides meaningful uncertainty estimates that guide sensor placement and decision making, offering practical value for cost-constrained transportation systems.

Abstract

Traffic prediction is a crucial topic because of its broad scope of applications in the transportation domain. Recently, various studies have achieved promising results. However, most studies assume the prediction locations have complete or at least partial historical records and cannot be extended to non-historical recorded locations. In real-life scenarios, the deployment of sensors could be limited due to budget limitations and installation availability, which makes most current models not applicable. Though few pieces of literature tried to impute traffic states at the missing locations, these methods need the data simultaneously observed at the locations with sensors, making them not applicable to prediction tasks. Another drawback is the lack of measurement of uncertainty in prediction, making prior works unsuitable for risk-sensitive tasks or involving decision-making. To fill the gap, inspired by the previous inductive graph neural network, this work proposed an uncertainty-aware framework with the ability to 1) extend prediction to missing locations with no historical records and significantly extend spatial coverage of prediction locations while reducing deployment of sensors and 2) generate probabilistic prediction with uncertainty quantification to help the management of risk and decision making in the down-stream tasks. Through extensive experiments on real-life datasets, the result shows our method achieved promising results on prediction tasks, and the uncertainty quantification gives consistent results which highly correlated with the locations with and without historical data. We also show that our model could help support sensor deployment tasks in the transportation field to achieve higher accuracy with a limited sensor deployment budget.

Figures (3)

• Figure 1: Framework of UIGNN. (a) The missing locations indexes are 7,10. And during training, we random sample sub-graph and mask locations 4,6 (upper sample) and train GNN to impute mask locations from time $t-T$ to $t$ and predict all sampled locations at $t+\Delta t$ and quantify the prediction uncertainty. (b) During prediction, UIGNN predicts both missing and observable locations at time $t+\Delta t$ and quantifies the prediction uncertainty.
• Figure 2: Model uncertainty of 30 mins task prediction task on different locations of the METR-LA road network. (a) and (b) shows RMSE and NLL on different locations of the road network, and the starts and dots mean missing and observable locations. Compared to observable locations, missing locations have higher error and model uncertainty. (c) shows the Epistemic uncertainty of predictions at different locations. Missing locations with more and closer observable neighbors have lower model uncertainty, and observable locations do not have this effect. (d) shows that the error increases with the epistemic uncertainty increasing in both observable and missing locations.
• Figure 3: Performance of Uncertainty sampling and random sampling method on active the sensing task. Lower RMSE means better prediction performance. (a) It shows different sampling methods have similar performance on observable locations. (b) The uncertainty sampling method performs better compared to the random sampling method. And two methods finally converge and achieve the same level of accuracy.