In brief

By refining the loss function, researchers demonstrate how mitigating inaccuracies in data overfitting can create more reliable machine learning platforms for traffic prediction and other applications.

© Unsplash

Traffic prediction framework gets the green light

16 Aug 2023

A new machine learning platform offers more accurate traffic predictions for safer, less congested roads.

It’s hard to escape bumper-to-bumper traffic during peak hours, but high-powered computational platforms can help us get our cities moving smoothly again. Traffic prediction analyses large amounts of data from traffic sensors and is an important aspect of managing traffic flow.

“Accurate traffic prediction empowers road users to make informed decisions and contributes to the alleviation of traffic congestion,” explained Peisheng Qian and Ziyuan Zhao, research engineers at A*STAR’s Institute for Infocomm Research (I2R). They added that these systems can also enhance the efficiency of public transport systems and promote safety.

However, the root of the challenge lies in the dynamic, ever-changing nature of traffic. Moreover, road congestion is dependent not only on the current road conditions, but also on events that may have already occurred. One machine learning model, spatial-temporal graph neural networks (GNNs), is particularly suited for this problem with its ability to make deductions using historical information from any given node alongside neighbouring nodes.

Nonetheless, GNNs face limitations when the target prediction is outside the primary region of interest in the data distribution, which is known to cause accuracy levels to plummet. Qian and Zhao worked with colleagues from Sichuan University, China and the Nanyang Technological University, Singapore to fundamentally change how the machine learning model is optimised for the accuracy of traffic predictions.

“Choosing an appropriate loss function prevents regression models from overfitting to outlier points, or classification models from overfitting to the majority class,” noted the duo.

The team’s new loss function, dubbed mean-residue loss, involves a two-step process: conditional distribution learning followed by speed regression. In the first step, traffic sensor data is analysed to understand the conditional probability distribution of speeds given certain conditions. Next, information from the learned conditional distribution is leveraged to predict the actual speed values.

They tested their model against current state-of-the-art traffic prediction platforms using three widely used traffic datasets, with the results of experimental and theoretical demonstrations putting their model in the top spot.

“Our approach can be advantageous in other domains characterised by spatial-temporal correlations, including bioinformatics, climate science, video analysis, and supply chain management,” concluded Qian, who said the team is currently exploring their GNN mean-residue loss approach for other applications, e.g., age estimation.

The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).

Want to stay up to date with breakthroughs from A*STAR? Follow us on Twitter and LinkedIn!


Zeng, Z., Zhao, W., Qian, P., Zhou, Y., Zhao, Z., et al. Robust traffic prediction from spatial-temporal data based on conditional distribution learning. IEEE Transactions on Cybernetics 52 (12), 13458-13471 (2022). | article

About the Researchers

Peisheng Qian received a BSc degree in computer science from the National University of Singapore in 2019. He is currently a Research Engineer at A*STAR’s Institute for Infocomm Research (I2R). His research interests include computer vision and deep learning.
Ziyuan Zhao is currently a Senior Research Engineer at A*STAR’s Institute for Infocomm Research (I2R). Previously, he received an MTech degree at the National University of Singapore in 2019 and a BEng degree at Yunnan University, China, in 2017. His research interests are in the fields of deep learning, machine learning, medical imaging, bioinformatics, healthcare and artificial intelligence.

This article was made for A*STAR Research by Wildtype Media Group