Scalable Learning of Segment-Level Traffic Congestion Functions

TL;DR

The paper addresses learning segment-level congestion relationships at global scale by pooling static and dynamic segment features and training a single data-driven congestion function (CF) per road type. It demonstrates that a neural CF trained on pooled data can predict segment speeds from partial flow across multiple cities, generalize to unseen segments, and transfer across cities, particularly for highways. It also shows the pooled model can infer segment properties such as critical density and discusses trade-offs versus per-segment, model-based baselines. The work offers a scalable approach for city-scale mesoscopic traffic modeling with potential impact on real-time navigation and policy analysis, while outlining directions for improvements like richer features and graph-based architectures.

Abstract

We propose and study a data-driven framework for identifying traffic congestion functions (numerical relationships between observations of traffic variables) at global scale and segment-level granularity. In contrast to methods that estimate a separate set of parameters for each roadway, ours learns a single black-box function over all roadways in a metropolitan area. First, we pool traffic data from all segments into one dataset, combining static attributes with dynamic time-dependent features. Second, we train a feed-forward neural network on this dataset, which we can then use on any segment in the area. We evaluate how well our framework identifies congestion functions on observed segments and how it generalizes to unobserved segments and predicts segment attributes on a large dataset covering multiple cities worldwide. For identification error on observed segments, our single data-driven congestion function compares favorably to segment-specific model-based functions on highway roads, but has room to improve on arterial roads. For generalization, our approach shows strong performance across cities and road types: both on unobserved segments in the same city and on zero-shot transfer learning between cities. Finally, for predicting segment attributes, we find that our approach can approximate critical densities for individual segments using their static properties.

Figures (5)

• Figure 1: An overview of our approach. We collect datapoints that capture static and dynamic properties of each segment within a given city. We then segregate datapoints by road type, in particular whether they are from a highway or arterial road. Finally, we train a separate neural network on each class of samples to estimate mean speed given current observed volume and other features. The overhead images are from Google Maps.
• Figure 2: The MAEs from \ref{['tab:aggregate_metrics']}, disaggregated by quartiles of normalized speed, i.e., speed divided by segment speed limit. We focus on highway segments of three representative cities here. Our approach has lower MAE in periods of free-flow and transition between congestion and free-flow, while the baseline is better during periods of most congestion.
• Figure 3: Our approach generalizes well to unseen segments. We train it on 80% of the segments for each dataset, and evaluate it on the remaining 20% for the same time period. We do this through 5-fold cross-validation and report the median MAPE of the 5 folds here (y-axis) against the MAPE of training and testing on the same set of segments (x-axis).
• Figure 4: For selected pairs of cities, we plot the MAE of training on the first city and testing on the second (y-axis), against the MAE of training and testing on the second (x-axis). For pairs where both cities are from USA or Europe, the points are closer to the x=y line, suggesting better transfer possibly due to similar traffic patterns. The opposite is true when one of cities is from Asia.
• Figure 5: Distributions of predicted critical densities from different methods.