No. 1046

Title:

Learning linear non-gaussian acyclic models for latent factors

Author(s):

Shimizu, Shohei (The Institute of Statistical Mathematics);
Hoyer, Patrik (University of Helsinki);
Hyvärinen, Aapo (University of Helsinki)

Key words:

Independent component analysis; structural equation models; causal analysis; latent factors

Abstract:

    Many algorithms have been proposed for discovery of causal relations among observed variables. However, one often wants to discover causal relations among latent factors (hidden variables) rather than observed variables. Some methods have been proposed for estimating linear acyclic models for latent factors that are linearly measured by observed variables. However, most of the methods make the gaussian assumption, and this leads to a number of indistinguishable models. In this paper, we propose that a non-gaussian assumption allows all the connection strength and structure of linear acyclic models for latent factors uniquely identified under standard assumptions. We also conduct experiments with artificial data to study the validity of our approach.