[Paper Review] Network-Scale Traffic Modeling and Forecasting with Graphical Lasso and Neural Networks

Shiliang Sun, Rongqing Huang+1 more|arXiv (Cornell University)|Dec 25, 2017

Gaussian Processes and Bayesian Inference5 references3 citations

TL;DR

This paper proposes a multi-link traffic flow forecasting framework combining Graphical Lasso (GL) for sparse dependency modeling and neural networks (NN), outperforming single-link and multi-task learning baselines. GL-NN achieves superior accuracy by extracting relevant historical flows across adjacent links, with experiments showing it consistently outperforms MMTL and other models on real-world data.

ABSTRACT

Traffic flow forecasting, especially the short-term case, is an important topic in intelligent transportation systems (ITS). This paper does a lot of research on network-scale modeling and forecasting of short-term traffic flows. Firstly, we propose the concepts of single-link and multi-link models of traffic flow forecasting. Secondly, we construct four prediction models by combining the two models with single-task learning and multi-task learning. The combination of the multi-link model and multi-task learning not only improves the experimental efficiency but also the prediction accuracy. Moreover, a new multi-link single-task approach that combines graphical lasso (GL) with neural network (NN) is proposed. GL provides a general methodology for solving problems involving lots of variables. Using L1 regularization, GL builds a sparse graphical model making use of the sparse inverse covariance matrix. In addition, Gaussian process regression (GPR) is a classic regression algorithm in Bayesian machine learning. Although there is wide research on GPR, there are few applications of GPR in traffic flow forecasting. In this paper, we apply GPR to traffic flow forecasting and show its potential. Through sufficient experiments, we compare all of the proposed approaches and make an overall assessment at last.

Motivation & Objective

• Address the limitation of single-link traffic flow forecasting by incorporating spatial dependencies from adjacent links.
• Improve prediction accuracy and efficiency by integrating multi-link modeling with machine learning techniques.
• Explore the application of Gaussian Process Regression (GPR) in traffic flow forecasting for uncertainty-aware predictions.
• Evaluate the effectiveness of Graphical Lasso (GL) in extracting relevant variables from large-scale traffic networks.
• Compare single-task and multi-task learning frameworks across different modeling architectures.

Proposed method

• Propose single-link and multi-link forecasting models to distinguish between link-specific and network-wide dependencies.
• Combine Graphical Lasso (GL) with neural networks (GL-NN) to build a sparse graphical model using L1-regularized inverse covariance matrix for feature selection.
• Apply multi-task learning (MTL) to jointly train predictions across multiple links, improving generalization and efficiency.
• Use backpropagation (BP) neural networks as the core regression model in GL-NN and MMTL frameworks.
• Integrate Gaussian Process Regression (GPR) as a single-link single-task baseline, leveraging its predictive distribution for uncertainty quantification.
• Use real-world traffic flow data to train and evaluate all models, with performance measured via mean absolute error (MAE) and t-tests for significance.

Experimental results

Research questions

• RQ1 Can multi-link modeling that incorporates adjacent link data improve short-term traffic flow forecasting accuracy compared to single-link models?
• RQ2 Does combining Graphical Lasso with neural networks (GL-NN) yield better performance than multi-task learning (MMTL) in network-scale traffic forecasting?
• RQ3 How does Gaussian Process Regression (GPR) compare to neural network-based models in terms of prediction accuracy and uncertainty estimation for traffic flow?
• RQ4 To what extent does the use of sparse graphical models via GL enhance feature selection and prediction efficiency in large-scale traffic networks?
• RQ5 Is multi-task learning more effective than single-task learning when applied to multi-link traffic forecasting?

Key findings

• The multi-link single-task approach GL-NN outperforms the multi-link multi-task learning (MMTL) model on 22 out of 31 links in real-world data, with 21 links showing statistically significant improvement.
• GL-NN achieves the highest prediction accuracy among all four proposed models (SSTL, SMTL, MSTL, MMTL), demonstrating the effectiveness of GL in extracting relevant dependencies from network-wide data.
• A t-test confirms that GL-NN significantly outperforms MMTL, validating the superiority of GL-based feature selection over joint multi-task optimization.
• Gaussian Process Regression (GPR) provides predictive distributions with uncertainty bounds, and reducing noise variance allows the prediction interval to fully contain actual values, highlighting its potential for robust forecasting.
• Multi-link approaches consistently outperform single-link models, and multi-task learning frameworks yield better results than single-task learning, confirming the value of leveraging spatial and temporal correlations.