Spatio-Temporal Recursive Method for Traffic Flow Interpolation

Gang Wang, Yuhao Mao, Xu Liu, Haohan Liang, Keqiang LiSymmetry 2025 · DOI

TL;DR

This paper introduces the Temporal-Spatial Fusion Neural Network (TSFNN), a novel deep learning framework designed to accurately interpolate missing traffic flow data by simultaneously modeling both temporal and spatial dependencies. By effectively combining these two crucial aspects, TSFNN achieves superior performance compared to existing methods, offering a more robust solution for maintaining the integrity of traffic monitoring systems.

What Problem Does This Paper Address?

Accurate and complete traffic flow data is essential for effective transportation management, operational efficiency, and public safety. However, real-world traffic data often suffers from missing values due to sensor failures, communication issues, or data processing errors. This data incompleteness poses a significant challenge for downstream analytics and decision-making. While various interpolation methods exist, they often struggle to fully capture the inherent spatio-temporal correlations present in traffic data. Traditional static methods ignore temporal dynamics, while many dynamic methods focus on either temporal or spatial aspects, but not both simultaneously. This paper addresses the critical need for a method that can effectively leverage both temporal and spatial information to achieve highly accurate traffic flow interpolation, especially in scenarios with continuous missing data segments.

Methods

The core of this paper is the Temporal-Spatial Fusion Neural Network (TSFNN). This architecture is designed to jointly model spatio-temporal dependencies. It comprises two main components:

  • Temporal Module: This module is responsible for capturing sequential dependencies within individual traffic sensor streams. It utilizes a bidirectional Long Short-Term Memory (Bi-LSTM) network. Bi-LSTMs are chosen because they can process sequences in both forward and backward directions, allowing them to better capture long-range temporal patterns and mitigate the vanishing gradient problem common in standard RNNs.
  • Spatial Module: This module focuses on modeling inter-sensor dependencies, i.e., how traffic patterns at one sensor relate to patterns at others. It employs a Multilayer Perceptron (MLP). Notably, the MLP incorporates a self-masking parameter to learn spatial information mining and implicitly captures spatial relationships without relying on explicit distance matrices or adjacency graphs.

These two modules are then fused together. The TSFNN uses a weighting factor to combine the temporal and spatial estimates, adaptively learning the optimal balance between them based on temporal gap indicators and masking vectors. The model is trained by minimizing the Mean Absolute Error (MAE) between the estimated and observed values, aiming for distributional consistency.

Datasets

The authors evaluated their TSFNN model on four widely used real-world traffic datasets:

  • PeMS03
  • PeMS04
  • PeMS07
  • PeMS08

These datasets were sourced from the Caltrans Performance Measurement System. To thoroughly test the model's robustness, the researchers introduced controlled missingness into these datasets, creating continuous temporal gaps and also employing a Missing Completely at Random (MCAR) scheme. This hybrid missing pattern allowed for a comprehensive evaluation of the model's performance under various missing data scenarios.

Key Insights and Ideas

  • The Synergy of Spatio-Temporal Modeling: The most significant insight is that effectively interpolating traffic data requires a simultaneous consideration of both temporal and spatial correlations. Neither temporal nor spatial modeling alone is sufficient for achieving high accuracy, especially under challenging missing data conditions.
  • Implicit Spatial Relationship Learning: The use of an MLP with self-masking parameters to learn spatial dependencies without explicit distance matrices or graphs is a clever approach. It suggests that complex spatial relationships can be implicitly learned from the data itself.
  • Bidirectional Temporal Modeling is Crucial: The ablation studies highlight that while unidirectional RNNs can capture some temporal patterns, the bidirectional nature of Bi-LSTM significantly enhances interpolation accuracy by considering dependencies from both past and future observations.
  • Robustness Across Missing Rates: TSFNN demonstrates consistent superiority across different missing data rates, indicating its reliability for real-world applications where missing data can vary significantly.

Key Results

The TSFNN model consistently outperformed all baseline methods across the four datasets and various missing rates, as evidenced by both MAE and RMSE metrics. For instance, in Table 1 (MAE), TSFNN achieved the lowest MAE values, such as 7.8933 for PEMS03 at 30% missing data, significantly outperforming other methods. Similarly, in Table 2 (RMSE), TSFNN also showed the best performance, with an RMSE of 14.4309 for PEMS03 at 30% missing data. The visual results in Figure 6 further illustrate TSFNN's ability to accurately reconstruct missing traffic flow patterns, with its interpolated curves closely aligning with the ground truth. The ablation studies confirmed that removing either the temporal or spatial module, or reducing the temporal module to unidirectional, led to a noticeable drop in performance, underscoring the indispensable role of each component.

Limitations and Future Work

The paper acknowledges that while TSFNN was tested on freeway-based PeMS datasets, its modular design is not inherently limited to such scenarios. The authors suggest that TSFNN could be adapted to other traffic environments like arterial roads and urban intersections. Furthermore, they propose that the framework has potential applicability in other domains involving spatio-temporal missing data, such as meteorological forecasting, environmental sensing, or healthcare monitoring. Future work will focus on exploring domain adaptation to broaden TSFNN's applicability.

Summary updated

BibTeX
@article{wang2025spatio,
  author = {Gang Wang and Yuhao Mao and Xu Liu and Haohan Liang and Keqiang Li},
  title = {{Spatio-Temporal Recursive Method for Traffic Flow Interpolation}},
  journal = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17091577},
}
Figure