Abstract
Accurate traffic prediction is essential for intelligent transportation systems, yet existing models often struggle with two critical real-world challenges: the absence of historical data for many road segments (the sensor gap) and the lack of reliable confidence metrics for predictions (the confidence gap). This paper introduces the Uncertainty-aware Inductive Graph Neural Network (UIGNN), a framework that leverages inductive graph learning and evidential deep regression to provide high-accuracy forecasts for unmonitored locations while simultaneously quantifying predictive uncertainty. Our results across three large-scale datasets demonstrate that UIGNN outperforms traditional "impute-then-predict" pipelines and provides uncertainty metrics that can effectively guide urban sensor deployment.
Introduction
Traffic forecasting powers modern urban infrastructure, from GPS routing to autonomous vehicle coordination. However, most state-of-the-art models are transductive, meaning they rely on the assumption that every location requiring a prediction has a dedicated sensor and a comprehensive historical record. In practice, budget constraints and physical limitations leave many road segments "dark," with no historical data available. Furthermore, most models provide only point estimates, offering no measure of certainty. For risk-sensitive applications like emergency response or traffic signal control, this lack of confidence intervals can lead to suboptimal or even dangerous decision-making.
To address these issues, we propose UIGNN, a single-step solution that generalizes to unseen locations while quantifying its own "known unknowns."
Methodology
The UIGNN Framework
The Uncertainty-aware Inductive Graph Neural Network (UIGNN) is built upon three primary pillars designed to handle spatial missingness and uncertainty:
- Inductive Graph Learning: Unlike traditional GNNs that learn fixed node embeddings, UIGNN learns to aggregate features from a node’s neighborhood. This allows the model to generalize to "unseen" nodes—locations without sensors—by leveraging the topological structure of the road network.
- Diffusion Graph Convolutional Networks (DGCN): To capture complex spatiotemporal dependencies, UIGNN utilizes DGCN as its backbone. This treats traffic flow as a diffusion process on a directed graph, effectively modeling how congestion propagates both forward and backward through the network.
- Evidential Deep Regression (EDL): To quantify uncertainty without the high computational cost of Bayesian methods (such as Monte Carlo Dropout), the model integrates EDL. This allows UIGNN to output a probability distribution in a single forward pass, helping to identify out-of-distribution (OOD) samples where traffic patterns are too unique for reliable prediction.
Training Strategy
During the training phase, the model employs a subgraph sampling and masking strategy. By randomly masking certain nodes, the model is forced to learn how to recover missing information and predict future states simultaneously. This integrated approach prevents the error propagation typically found in multi-stage "impute then predict" workflows.
Experimental Evaluation
Datasets
The model was evaluated on three real-world traffic speed datasets:
- METR-LA: 207 loop detectors on the Los Angeles highway network.
- PeMS-BAY: 325 sensors in the San Francisco Bay Area.
- SeData: 323 locations on a highway network with a binary adjacency matrix.
Performance and Reliability
UIGNN consistently achieved top-tier performance across all datasets. Notably, on the "missing" locations within the METR-LA dataset, UIGNN achieved an RMSE of 7.3052, a significant improvement over the IGNNK+DGCN baseline of 7.7083.
Beyond accuracy, the model's uncertainty quantification proved to be a reliable proxy for error. In locations without sensors where ground truth is unavailable, the Negative Log-Likelihood (NLL) reported by the model correlated strongly with actual error rates, providing a vital "health check" for predictions in unmonitored areas.
Practical Applications: Guided Sensor Deployment
A key advantage of UIGNN is its ability to inform urban planning. In an "Active Sensing" case study, the authors demonstrated that using the model’s uncertainty metrics to select locations for new sensors resulted in a much faster reduction in overall prediction error compared to random sensor placement. This suggests that city planners can use UIGNN to maximize the "information profit" of limited infrastructure budgets.
Conclusion and Future Work
UIGNN represents a significant step toward robust traffic prediction in data-sparse environments. By combining inductive learning with evidential regression, it bridges the gap between high-performance forecasting and practical reliability.
Future research will focus on expanding the model to handle dynamic topologies (such as road closures), optimizing the subgraph sampling strategy for city-wide scaling, and addressing temporal missingness—specifically "flickering" sensors that provide intermittent data rather than being entirely absent.
Summary updated