Pot-luck challenge: FACT SHEET

We designed four datasets for the purpose of benchmarking local causal discovery algorithms. These include two “re-simulated” datasets obtained from artificially generated data from models trained with real data and two datasets including real variables intermixed with artificial variables (called probes). There is no time dependency in the samples. We chose applications in marketing, pharmacology and bio-medicine spanning a high diversity of types of distributions. The datasets were used in two challenges in 2008 organized for the WCCI and NIPS conferences. A detailed technical report on the dataset design is available (Guyon et al., 2009). The website of the challenges remains open for post-challenge submissions (http://clopinet.com/causality).