A causal understanding of climatic interactions is of high societal relevance from identifying causes of extreme events to process understanding and weather forecasting. Classically, causal relations in climate research are investigated using climate models that simulate the interactions of the atmosphere, water bodies, land surface and the cryosphere. However, these are very expensive, time-consuming, and cannot resolve all physical processes. Fortunately, recent decades have seen an explosion in the availability of large-scale time series data (satellite remote sensing, station-based, or field site measurements). Such data repositories, together with increasing computational power, open up novel ways to use data-driven methods for observational causal discovery. Observational causal inference is a major current topic in machine learning as well as other domains from physics to applied fields. However, despite rapid methodological progress, many methods are not widely known in applied fields such as Earth sciences. A major impediment that we aim to overcome with this competition is the lack of reliable benchmarks on how methods deal with various real world challenges. This NeurIPS competition comprises a number of multivariate time series datasets featuring major challenges of climate data from time delays and nonlinearity to nonstationarity and selection bias. We aim to open up new interdisciplinary research pathways by spurring research on novel causal discovery methods for climate data analysis.
Prize Money and AWS computation credits: Thanks to a very generous sponsoring by Amazon AWS, we can offer a total of USD $10,000 in prize money which will be distributed in prizes for different categories and an overall prize as follows:
$250 for the best ranking participant for a particular dataset
$1,500 for the best ranking participant across all datasets
A participant can earn awards in multiple categories. During the competition, we provide AWS computation credits to enable also users with less computational resources to participate.
The dates for this competition are as follows:
Start of testing phase: July 31, 2019
In this phase we will provide datasets on a regular basis, starting from easy to very challenging datasets that are similar to the final evaluation datasets.
Start of main competition: Oct 11, 2019
In this phase we will provide challenging climate and weather datasets in different categories on which the final ranking will be evaluated.
End of competition: Oct 31, 2019
NeurIPS competition session (Vancouver, Canada): Dec 13, 2019
This competition finishes in Dec 2019, you can continue benchmarking on the main CauseMe platform
To get an overview over the platform for the C4C competition, it's best to signup for the main CauseMe platform and watch the video tutorial in the HowTo page there.
C4C covers a wide range of datasets featuring a number of climate and weather data challenges (see figure). Participants download datasets, run their own causal discovery method, and upload their predictions (matrices of causal connections). The platform evaluates and ranks the methods according to the AUC performance metric. After signing up and logging in, more information and example code snippets are given.
If you find the Causality 4 Climate competition useful, acknowledge it by citing:
Inferring causation from time series in Earth system sciences. J. Runge, S. Bathiany, E. Bollt, G. Camps-Valls, D. Coumou, E. Deyle, C. Glymour, M. Kretschmer, M.D. Mahecha, J. Munoz-Mari, E.H. van Ness, J. Peters, R. Quax, M. Reichstein, M. Scheffer, B. Schölkopf, P. Spirtes, G. Sugihara, J. Sun, K. Zhang, J. Zscheischler. Nature Communications 10: 2553 (2019).