Traffic Map Movie Forecasting 2021
- Study high-resolution 8-channel traffic movies of entire cities together with road network information.
- Predict future traffic flows across spatial (unseen city) and temporal (pre- and post-COVID-19) domain shifts
- Discover underlying rules and patterns in movies or on graphs.
Building on its success at NeurIPS 2019 and 2020, the Traffic4cast competition is going into its third year offering new challenges and opportunities. Our Traffic Map Movies have been built from 1012 data points, an order of magnitude more than last year, covering ten cities across 2019 and 2020. This gives us the flexibility to let participants model urban traffic, which is governed by largely unknown, implicit rules, to predict future states robustly even in the face of complex domain shifts. We explore this generalizability through few-shot learning challenges of growing complexity. Coping with a drastic temporal domain shift, models need to contend with the changing traffic patterns induced by the COVID-19 pandemic. In an extended challenge, predictions are requested for new cities never seen before. Only a few traffic measurements are provided, supported by a summarial characterization of the road network. The competition thus brings together a range of highly active fields of machine learning — few-shot learning, transfer learning, and meta-learning more generally, as well as video frame prediction or graph based modelling.
Notably, we have seen that insights from our t4c competitions can be applied to other spatial processes, such as weather. Results will therefore be of direct interest not only to applied research in traffic or weather forecasting but also to models of other spatial processes (spread of pandemics or wildfires, evolving land use, etc., contributing to jointly addressing some of humanities most pressing problems together.
Insights from our previous competitions
In a nutshell,
- Our chosen representation was highly effective and — as independent work has shown for precipitation prediction — should be considered a promising new technique for tackling complex geo-spatial processes.
- Our 2020 competition asked participants to predict traffic in the second half of 2019 from given data in the first half of 2019. Thus temporal transfer learning across seasons was possible, although a performance boost could be achieved when also using the validation set data provided for training.
- The work of H. Martin following their 2019 prize winning solution indicated that some degree of transfer of learned traffic patterns to almost unseen cities was possible with their GNN approach.
Now, going beyond the challenges at NeurIPS 2019 and 2020, this year will explore models that adapt to domain shift both in space and time. Specifically, we will provide dynamic traffic data for 4 different cities in the more effective format we introduced in Traffic4cast 2020. Half of this data will be from 2019, before the COVID pandemic hit the world, and the other half will be from 2020 when the pandemic started affecting all our lives. We will complement these dynamic data by static information on road geometry. We then provide two challenges to participants:
- In the core challenge, participants are tasked to handle temporal domain shift (an active field of machine learning research) in traffic due to COVID-19. In addition to the full data for the four cities described above, participants will receive pre-COVID data for four further cities, plus one hundred 1h slots from 2020 after COVID struck. The challenge then is to predict the dynamic traffic states 5, 10, 15, 30, 45 and 60 minutes into the future after each of the one hundred time slots for each of the additional 4 cities.
- In an extended challenge, participants are asked to again predict dynamic traffic states for two further cities, hitherto entirely unseen. Like for the first challenge, traffic needs to be predicted 5, 10, 15, 30, 45 and 60 minutes into the future following 100 given 1h time slots. Yet there will be no further traffic data provided for these cities. Moreover, for each city, 50 of these 100 1h time slots will be from the period before COVID, and 50 from the period after COVID, without revealing which!
A solution to the extended challenge could also be applied to the core challenge, thus truly making it more universal. Both challenges are few-shot learning challenges, an actively developing area of AI, and can be tackled in a multitude of ways given the data provided.
The common underlying scientific question is how to build robust models that can predict how a complex spatial process evolves over time, so that the models can swiftly adjust to domain shifts both in space and time after seeing only a few examples from the new regime.
24 May 2021: Announcement and advertising of the beta-test phase of the competition, community discussion forum updates for Traffic4cast 2021 go live.
15 June 2021: Release of traffic data for Traffic4cast 2021