We are excited to announce the opening of our Traffic4cast competition! For the fourth consecutive year, Traffic4cast got selected to be presented in the competition track at the NeurIPS conference. The competition data will be available for download on July 27, 2022.
The Taffic4cast competition series focuses on developing advanced machine learning algorithms for modeling complex traffic processes using unique datasets. Traffic4cast 2022 connects to and builds on our previous competitions in 2019, 2020, and 2021 to offer challenges on a new level.
This year, our dynamic graph data combine information from road maps, GPS trajectories, and traffic loop counters in three diverse metropolitan areas (London, Madrid, and Melbourne) for two years (2019 to 2021). The GPS trajectories from a large fleet of probe vehicles are provided by HERE Technologies. Unlike previous years, these data serve to obtain ground truth labels. Simplified road maps are used to derive directed graphs with edges and nodes corresponding to road segments and intersections, respectively. Publicly available traffic loop counter data are a new data source for this year’s competition. Loop counters are spatially sparse; their measurements, aggregated over 15 minute intervals, are mapped to graph nodes.
The competition comprises the core and extended challenges. In the core challenge, participants are asked to predict the congestion classes (red/yellow/green) for the entire road graph 15 minutes into the future. In the extended challenge, participants are invited to predict the average speeds on each road segment of the graph 15 minutes into the future.
Traffic forecasting is important for planning urban traffic and optimizing the use of urban space, which has a direct impact on our environment and society. More generally, predicting temporal edge features in graphs from sparse node data has various real-world applications, such as malware detection, modeling of network traffic or pandemic spread. The competition brings together researchers from a number of highly active areas of machine learning, including graph-based modeling, transfer learning, deep learning, and time series prediction.
Interested in advancing machine learning algorithms and understanding traffic in a new way?
Join the competition!