Introduction
Intelligent Transportation Systems (ITS) are fundamental to modern urban management, providing the data necessary for real-time navigation and long-term infrastructure planning. However, the reliability of these systems is often compromised by sensor malfunctions and communication failures, leading to missing data rates as high as 25%. Traditional imputation methods often fail to capture the complex, non-linear spatiotemporal dependencies inherent in traffic networks. This paper introduces STAWnet (Spatio-Temporal Attention WaveNet), a framework designed to reconstruct missing traffic data by leveraging both historical and future context alongside dynamic spatial relationships.
Methodology: The STAWnet Framework
STAWnet adapts advanced architectures from traffic forecasting to the specific requirements of data imputation. Unlike prediction, which is purely autoregressive, imputation benefits from "bidirectional" context—using observations from both before and after a gap to estimate missing values.
The architecture consists of several innovative components:
- Gated Temporal Convolutional Networks (TCN): By utilizing dilated convolutions, STAWnet achieves a vast receptive field without the sequential processing bottlenecks of Recurrent Neural Networks (RNNs). This allows the model to capture long-range temporal patterns essential for understanding daily and weekly traffic cycles.
- Dynamic Attention Network (DAN): Rather than assuming static relationships between sensors based on physical distance, the DAN calculates "functional" distance in real-time. It dynamically weights the influence of neighboring sensors at each time step, accounting for shifting traffic conditions like congestion or accidents.
- Self-Learning Node Embeddings: The model automatically learns high-dimensional representations (embeddings) for each sensor. These embeddings capture the intrinsic characteristics of specific road segments (e.g., highway vs. residential street) without requiring manual feature engineering.
Experimental Evaluation
The framework was validated using three diverse real-world datasets: the GZ dataset (traffic speed from 214 road segments in Guangzhou), the HZ dataset (passenger flow from 80 metro stations in Hangzhou), and the PeMS dataset (traffic speeds from 325 sensors in the California Bay Area).
Results and Analysis
STAWnet was compared against two established benchmarks: Bayesian Temporal Matrix Factorization (BTMF) and Low-Rank Tensor Completion (LRTC-TNN).
As shown in the results table, STAWnet demonstrates significant performance gains, particularly on larger datasets. On the PeMS dataset with a 20% missing rate, STAWnet achieved a Mean Absolute Error (MAE) of 3.36, a dramatic improvement over the benchmarks which exceeded 58.0. Similarly, on the GZ dataset, STAWnet consistently maintained the lowest error rates across all metrics (MAE, MAPE, and RMSE) for both 20% and 40% missing data scenarios.
However, the results also highlight a "data hunger" characteristic of deep learning models. In the HZ metro dataset—the smallest of the three—STAWnet was outperformed by the LRTC-TNN method. This suggests that while STAWnet is highly effective for large-scale networks, its performance is contingent on having sufficient data to properly train the node embeddings and attention mechanisms.
Conclusion
STAWnet represents a significant advancement in traffic data imputation by moving beyond static graph assumptions and leveraging bidirectional temporal context. While the model requires substantial data to reach peak performance, its ability to capture dynamic spatiotemporal dependencies makes it a powerful tool for maintaining the data integrity of modern Intelligent Transportation Systems. Future work will focus on improving the model's efficiency on smaller datasets and exploring the interpretability of its learned attention weights.
Summary updated