Traffic4cast 2020 -- Graph Ensemble Net and the Importance of Feature And Loss Function Design for Traffic Prediction

TL;DR

This paper presents a traffic prediction solution for Traffic4cast 2020, emphasizing feature and loss function design, and introduces a novel ensemble GNN architecture that, combined with U-Nets, achieved 4th place.

Contribution

It introduces a new ensemble GNN architecture and analyzes the impact of feature and loss function design on traffic prediction accuracy.

Findings

Ensemble of U-Net and GNN achieved top performance

Feature and loss function design significantly affect results

GNN outperformed previous models but lagged behind U-Nets

Abstract

This paper details our solution to Traffic4cast 2020. Similar to Traffic4cast 2019, Traffic4cast 2020 challenged its contestants to develop algorithms that can predict the future traffic states of big cities. Our team tackled this challenge on two fronts. We studied the importance of feature and loss function design, and achieved significant improvement to the best performing U-Net solution from last year. We also explored the use of Graph Neural Networks and introduced a novel ensemble GNN architecture which outperformed the GNN solution from last year. While our GNN was improved, it was still unable to match the performance of U-Nets and the potential reasons for this shortfall were discussed. Our final solution, an ensemble of our U-Net and GNN, achieved the 4th place solution in Traffic4cast 2020.