Causality Discovery with Additive Disturbances: An Information-Theoretical Perspective

  • Authors:
  • Kun Zhang;Aapo Hyvärinen

  • Affiliations:
  • Dept of Computer Science & HIIT, University of Helsinki, Finland;Dept of Computer Science & HIIT, University of Helsinki, Finland and Dept of Mathematics and Statistics, University of Helsinki, Finland

  • Venue:
  • ECML PKDD '09 Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases: Part II
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

We consider causally sufficient acyclic causal models in which the relationship among the variables is nonlinear while disturbances have linear effects, and show that three principles, namely, the causal Markov condition (together with the independence between each disturbance and the corresponding parents), minimum disturbance entropy, and mutual independence of the disturbances, are equivalent. This motivates new and more efficient methods for some causal discovery problems. In particular, we propose to use multichannel blind deconvolution, an extension of independent component analysis, to do Granger causality analysis with instantaneous effects. This approach gives more accurate estimates of the parameters and can easily incorporate sparsity constraints. For additive disturbance-based nonlinear causal discovery, we first make use of the conditional independence relationships to obtain the equivalence class; undetermined causal directions are then found by nonlinear regression and pairwise independence tests. This avoids the brute-force search and greatly reduces the computational load.