Non-linear Causal Inference using Gaussianity Measures

We provide theoretical and empirical evidence for a type of asymmetry between causes and e ects that is present when these are related via linear models contaminated with additive non-Gaussian noise. Assuming that the causes and the e ects have the same distribution, we show that the distribution of...

Descripción completa

Detalles Bibliográficos
Autores: Hernández Lobato, Daniel, Morales Mombiela, Pablo, López-Paz, David, Suárez González, Alberto
Tipo de recurso: artículo
Fecha de publicación:2016
País:España
Institución:Universidad Autónoma de Madrid
Repositorio:Biblos-e Archivo. Repositorio Institucional de la UAM
Idioma:inglés
OAI Identifier:oai:repositorio.uam.es:10486/675246
Acceso en línea:http://hdl.handle.net/10486/675246
Access Level:acceso abierto
Palabra clave:Causal inference
Gaussianity of the residuals
Cause-effect pairs
Informática
Descripción
Sumario:We provide theoretical and empirical evidence for a type of asymmetry between causes and e ects that is present when these are related via linear models contaminated with additive non-Gaussian noise. Assuming that the causes and the e ects have the same distribution, we show that the distribution of the residuals of a linear t in the anti-causal direction is closer to a Gaussian than the distribution of the residuals in the causal direction. This Gaussianization e ect is characterized by reduction of the magnitude of the high-order cumulants and by an increment of the di erential entropy of the residuals. The problem of non-linear causal inference is addressed by performing an embedding in an expanded feature space, in which the relation between causes and e ects can be assumed to be linear. The e ectiveness of a method to discriminate between causes and e ects based on this type of asymmetry is illustrated in a variety of experiments using di erent measures of Gaussianity. The proposed method is shown to be competitive with state-of-the-art techniques for causal inference.