[Paper Review] Deep Learning Predictors for Traffic Flows

Nicholas Polson, Vadim Sokolov|arXiv (Cornell University)|Apr 15, 2016

Traffic Prediction and Management Techniques20 citations

TL;DR

This paper proposes a deep learning model combining ℓ₁-regularized linear layers with ReLU activation layers to predict traffic flows, capturing sharp nonlinear transitions between traffic regimes. It demonstrates superior short-term prediction accuracy during sudden events like football games and snowstorms on I-55.

ABSTRACT

We develop a deep learning model to predict traffic flows. The main contribution is development of an architecture that combines a linear model that is fitted using $\ell_1$ regularization and a sequence of $ anh$ layers. The challenge of predicting traffic flows are the sharp nonlinearities due to transitions between free flow, breakdown, recovery and congestion. We show that deep learning architectures can capture these nonlinear spatio-temporal effects. The first layer identifies spatio-temporal relations among predictors and other layers model nonlinear relations. We illustrate our methodology on road sensor data from Interstate I-55 and predict traffic flows during two special events; a Chicago Bears football game and an extreme snowstorm event. Both cases have sharp traffic flow regime changes, occurring very suddenly, and we show how deep learning provides precise short term traffic flow predictions.

Motivation & Objective

• To address the challenge of predicting abrupt traffic regime changes such as breakdowns and congestion.
• To develop a deep learning architecture capable of modeling complex spatio-temporal nonlinearities in traffic flow.
• To improve short-term traffic flow prediction accuracy during extreme events like snowstorms and major sporting events.
• To integrate linear spatio-temporal dependencies with nonlinear transformation through stacked ReLU layers.

Proposed method

• The model uses a linear layer with ℓ₁ regularization to identify key spatio-temporal predictors.
• Subsequent layers apply ReLU (‘anh’) activation functions to model nonlinear relationships in traffic dynamics.
• The architecture captures hierarchical dependencies: early layers detect spatial and temporal patterns, later layers model complex nonlinear transitions.
• The model is trained on road sensor data from Interstate I-55 to predict short-term traffic flow.
• The combination of sparse linear modeling and deep nonlinear transformation enables robust handling of regime shifts.
• The framework is evaluated on two real-world events with abrupt flow changes: a Chicago Bears game and a snowstorm.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively predict traffic flow during sudden regime shifts like congestion onset?
• RQ2 How well does the integration of ℓ₁-regularized linear layers and ReLU activation layers capture spatio-temporal nonlinearities in traffic data?
• RQ3 Does the proposed architecture outperform traditional models in short-term traffic prediction during extreme events?
• RQ4 To what extent can the model detect and respond to sharp transitions between free flow, breakdown, and recovery states?

Key findings

• The model successfully captures sharp nonlinear transitions in traffic flow during both the football game and snowstorm events.
• The inclusion of ℓ₁ regularization improves feature selection and model interpretability in high-dimensional traffic predictor spaces.
• The deep architecture with ReLU layers provides more accurate short-term predictions than baseline models.
• The model demonstrates robustness in predicting rapid regime changes, such as sudden congestion onset.
• The first layer effectively identifies relevant spatio-temporal relationships among traffic sensors.
• The sequential structure enables progressive modeling of complex traffic dynamics from linear to nonlinear patterns.