[Paper Review] Incrementally Improving Graph WaveNet Performance on Traffic Prediction

Sam Shleifer, Clara H. McCreery+1 more|arXiv (Cornell University)|Dec 11, 2019

Traffic Prediction and Management Techniques4 references17 citations

TL;DR

This paper incrementally improves Graph WaveNet's performance on traffic prediction by optimizing hyperparameters, adding skip connections for better gradient flow, and pretraining on short-term forecasting. These changes reduce Mean MAE by 0.054 on METR-LA and 0.055 on PEMS-BAY, achieving state-of-the-art results with minimal architectural changes.

ABSTRACT

We present a series of modifications which improve upon Graph WaveNet's previously state-of-the-art performance on the METR-LA traffic prediction task. The goal of this task is to predict the future speed of traffic at each sensor in a network using the past hour of sensor readings. Graph WaveNet (GWN) is a spatio-temporal graph neural network which interleaves graph convolution to aggregate information from nearby sensors and dilated convolutions to aggregate information from the past. We improve GWN by (1) using better hyperparameters, (2) adding connections that allow larger gradients to flow back to the early convolutional layers, and (3) pretraining on an easier short-term traffic prediction task. These modifications reduce the mean absolute error by .06 on the METR-LA task, nearly equal to GWN's improvement over its predecessor. These improvements generalize to the PEMS-BAY dataset, with similar relative magnitude. We also show that ensembling separate models for short-and long-term predictions further improves performance. Code is available at https://github.com/sshleifer/Graph-WaveNet .

Motivation & Objective

• To incrementally enhance Graph WaveNet's performance on traffic speed prediction without major architectural overhaul.
• To investigate whether small modifications—such as hyperparameter tuning, skip connections, and pretraining—can yield significant performance gains.
• To evaluate the impact of different input representations, including handling zero-speed readings, on model accuracy.
• To determine whether ensembling short- and long-term prediction heads improves overall performance.
• To understand the relative importance of temporal history, graph convolution, and learned adjacency matrices in spatio-temporal modeling.

Proposed method

• Increased the number of channels in convolutional layers from 32 to 40 to improve representational capacity.
• Added skip connections that allow gradients to flow more effectively to early convolutional layers, improving training stability.
• Replaced zero-speed values in input features with the average speed across all sensors to better represent missing data.
• Applied learning rate decay during training to stabilize convergence and improve generalization.
• Pretrained the model on a short-term traffic prediction task (15-minute horizon) before fine-tuning on the full 60-minute prediction task.
• Ensembled separate models trained for short-term (15 min) and long-term (60 min) prediction to improve overall accuracy.

Experimental results

Research questions

• RQ1 Can small, incremental modifications to Graph WaveNet significantly improve its performance on traffic prediction tasks?
• RQ2 Does pretraining on a simpler, short-term forecasting task improve performance on the full 60-minute prediction task?
• RQ3 How do different input representations—especially handling zero-speed readings—affect model accuracy?
• RQ4 Is there a performance benefit to ensembling models trained for different prediction horizons?
• RQ5 What is the relative contribution of graph convolution and learned adjacency matrices to model performance?

Key findings

• The final model achieved a Mean MAE of 3.002 on the METR-LA dataset, a 0.055 reduction from the baseline Graph WaveNet (3.057).
• On the PEMS-BAY dataset, the improved model reduced Mean MAE by 0.055, demonstrating generalization beyond the original dataset.
• Removing any of the key modifications—such as learning rate decay or skip connections—resulted in a measurable performance drop, confirming their necessity.
• Pretraining on short-term forecasting improved full-task performance, though the fine-tuned model lost some accuracy on short-term horizons.
• The model's performance plateaued after using 5–6 timesteps (30 minutes) of history, indicating diminishing returns beyond that point.
• Without graph convolution, the model's Mean MAE rose to 3.59, and without the learned adjacency matrix, it reached 3.08, confirming the importance of both components.