We show how statistical compressions schemes, concretely the formalization of the Occam razor by the minimum description length (MDL), provide causal inference with a high precision in temporal and generally non-temporal data. We further prove an identifiability theorem for an MDL-based bivariate causal score for non-temporal data.

Bio: K. Schindlerova is a senior scientist and docent in the Data Mining and Machine Learning research group at the Faculty of Computer Science, University of Vienna, Austria. Her expertise spans neural network approximation theory, machine learning methods and causal inference, with a particular focus on temporal data. Schindlerova is the first or second author of more than 90 publications, primarily on causal inference and causal discovery, applied machine learning, and artificial neural networks. Her recent contribution includes introducing statistical compression schemes into Granger causal models to enable more efficient inference.