Urban traffic flow prediction: a spatio‐temporal variable selection‐based approach

Yanyan Xu, Hui Chen, Qing‐Jie Kong, Zhai Xi, Yuncai LiuJournal of Advanced Transportation 2015 · DOI

TL;DR

This paper introduces a novel approach for predicting urban traffic flow by leveraging information from all road segments in a network, not just adjacent ones. Their method, VS-SVR, uses a two-stage process: first, Multivariate Adaptive Regression Splines (MARS) intelligently selects the most relevant traffic data (spatio-temporal variables) from a high-dimensional input, and second, Support Vector Regression (SVR) uses these selected variables to make predictions. This approach significantly improves prediction accuracy compared to existing methods.

What Problem Does This Paper Address?

Accurate short-term traffic flow prediction is crucial for intelligent transportation systems (ITS), helping to manage congestion and provide better traveler information. Traditional methods often focus on the relationship between a target road segment and its immediate upstream and downstream neighbors. However, traffic conditions in a dense urban network are influenced by a much wider set of road segments. This creates a challenge: incorporating data from all segments leads to a high-dimensional input space, which can overwhelm prediction models. The paper aims to address this by developing a method that can effectively utilize this rich, high-dimensional spatio-temporal data while maintaining prediction accuracy.

Methods

The proposed method, VS-SVR (Variable Selection-based Support Vector Regression), operates in two stages:

  1. Spatio-temporal Variable Selection (using MARS):
    • Instead of just using adjacent road segments, the model considers traffic data from all available road segments in a given network area. This creates a high-dimensional set of input variables.
  2. Multivariate Adaptive Regression Splines (MARS) is employed to filter these variables. MARS is a non-linear regression technique that inherently performs variable selection by identifying the most influential predictors.
  3. It uses a two-stage process: a forward growing stage to build an over-fitted model and a backward pruning stage to remove less important basis functions.
  4. The output of this stage is a set of selected predictors with associated weights, effectively reducing the dimensionality of the input data while retaining the most relevant spatio-temporal information.
  5. Prediction (using SVR):
    • Support Vector Regression (SVR) is then trained using the weighted and selected spatio-temporal variables identified by MARS.
  6. SVR is chosen for its ability to handle non-linear relationships and its generalization capabilities.
  7. The paper uses an ε-insensitivity loss function, which makes SVR more tolerant to noise and robust.

This two-stage approach allows the model to harness the predictive power of a broad range of traffic data without being overwhelmed by its dimensionality.

Datasets

The study uses real-world traffic volume data collected from a sub-area of the Shanghai urban road network.

  • Data Source: Sydney Co-ordinated Adaptive Traffic System data, provided by Shanghai Traffic Information Center.
  • Data Granularity: Vehicle counts per 10-minute intervals, converted to vehicles per hour (veh/h).
  • Time Period: February 1-29, 2012 (excluding February 24).
  • Focus: Weekday traffic data, as weekday traffic is considered more complex and prone to congestion.
  • Network Size: A sub-area with 12 bidirectional road segments (24 individual road segments). Some road segments had unavailable data and were excluded from the analysis.
  • Training/Testing Split: The data was split into a training set (early 17 weekdays) and a testing set (remaining three weekdays).

Key Insights and Ideas

  • Beyond Adjacency: The most significant insight is that traffic flow at a target road segment is influenced by a much wider network than just its immediate neighbors. Considering all available road segments, even non-adjacent ones, can lead to more accurate predictions.
  • Dimensionality Reduction is Key: High-dimensional spatio-temporal data presents a significant challenge. The paper demonstrates that effective variable selection is crucial for making this rich data usable by prediction models like SVR.
  • MARS as a Smart Filter: MARS proves to be an effective tool for identifying relevant spatio-temporal predictors, acting as a "filter" that prepares the data for a powerful regression model like SVR.
  • SVR Benefits from Pre-selection: While SVR is a capable model, its performance on high-dimensional data without prior selection is suboptimal. The VS-SVR approach shows that pre-selecting variables significantly enhances SVR's predictive power.
  • Spatial Information is Crucial: The study highlights that spatio-temporal correlation approaches (like VS-SVR, ST-BMARS) generally outperform purely temporal models (like AR, SARIMA) in complex urban environments, underscoring the importance of spatial context.

Key Results

The proposed VS-SVR model consistently outperformed several benchmark models, including AR, MARS, SVR, SARIMA, and ST-BMARS, across various metrics (RMSE and MAPE) on the Shanghai traffic data.

  • Accuracy Improvement: For road I₁, VS-SVR achieved an RMSE of 71.58 and MAPE of 11.04, significantly lower than other models. For road D₂, it achieved an RMSE of 73.07 and MAPE of 12.34, also outperforming others.
  • Comparison Highlights:
    • Time-series models (AR, SARIMA) performed poorly, especially when traffic states changed suddenly.
  • MARS and SVR alone showed improvement but were surpassed by VS-SVR, indicating the benefit of the combined approach.
  • ST-BMARS, another spatio-temporal model, performed comparably to SVR but was still outperformed by VS-SVR.
  • Variable Selection Effectiveness: The variable selection phase identified relevant upstream, downstream, and even non-adjacent road segments that influenced the target road's traffic flow. For example, for road I₁, 10 out of 17 segments were selected.
  • Computational Cost: While the training phase of VS-SVR was more time-consuming due to variable selection, the prediction phase was efficient and suitable for real-time applications. The variable selection itself can be performed offline.

Limitations and Future Work

  • Model Sensitivity: The paper discusses the sensitivity of the model to parameters like the GCV penalty (μ) and the number of selected predictors (Np). They found that for road I₁, prediction error stabilized when μ ≥ 3 and when the number of selected variables was between 13 and 20.
  • Collinearity: While MARS can select collinear predictors, the study found that removing highly collinear variables did not significantly degrade performance, suggesting minor impact on the final accuracy.
  • Scalability: The authors acknowledge that while the model was tested on a small network, it can be extended to larger networks. For extremely large networks, a strategy of dividing the network into smaller sub-areas for processing is suggested.
  • Future Directions:
    • Investigating spatio-temporal correlations under different traffic states (e.g., morning peak, evening peak, stable traffic).
  • Finding an optimal balance between the road network scale and prediction accuracy.
  • Developing a robust road network partitioning model for large-scale traffic prediction.

Summary updated

Abstract

Summary Short‐term traffic flow prediction in urban area remains a difficult yet important problem in intelligent transportation systems. Current spatio‐temporal‐based urban traffic flow prediction techniques trend aims to discover the relationship between adjacent upstream and downstream road segments using specific models, while in this paper, we advocate to exploit the spatial and temporal information from all available road segments in a partial road network. However, the available traffic states can be high dimensional for high‐density road networks. Therefore, we propose a spatio‐temporal variable selection‐based support vector regression (VS‐SVR) model fed with the high‐dimensional traffic data collected from all available road segments. Our prediction model can be presented as a two‐stage framework. In the first stage, we employ the multivariate adaptive regression splines model to select a set of predictors most related to the target one from the high‐dimensional spatio‐temporal variables, and different weights are assigned to the selected predictors. In the second stage, the kernel learning method, support vector regression, is trained on the weighted variables. The experimental results on the real‐world traffic volume collected from a sub‐area of Shanghai, China, demonstrate that the proposed spatio‐temporal VS‐SVR model outperforms the state‐of‐the‐art. Copyright © 2015 John Wiley & Sons, Ltd.
BibTeX
@article{xu2015urban,
  author = {Yanyan Xu and Hui Chen and Qing‐Jie Kong and Zhai Xi and Yuncai Liu},
  title = {{Urban traffic flow prediction: a spatio‐temporal variable selection‐based approach}},
  journal = {Journal of Advanced Transportation},
  year = {2015},
  doi = {10.1002/atr.1356},
}
Figure