We suggest a novel strategy to linear causal discovery within the framework of multi-view Structural Equation Fashions (SEM). Our proposed mannequin relaxes the well-known assumption of non-Gaussian disturbances by alternatively assuming variety of variances over views, making it extra broadly relevant. We show the identifiability of all of the parameters of the mannequin with none additional assumptions on the construction of the SEM apart from it being acyclic. We additional suggest an estimation algorithm based mostly on latest advances in multi-view Unbiased Part Evaluation (ICA). The proposed methodology is validated by means of simulations and software on actual neuroimaging information, the place it permits the estimation of causal graphs between mind areas.
- † Inria, CEA, College of Paris-Saclay, France
- ‡ ENSAE, CREST, IP Paris, France
- § College of Helsinki, Finland
- * Equal Contributor
