Causal Discovery with Continuous Additive Noise Models

J. Peters; J.M. Mooij; D. Janzing; B. Schölkopf

Causal Discovery with Continuous Additive Noise Models

J. Peters, J.M. Mooij, D. Janzing, B. Schölkopf

114 Citations (Scopus)

Abstract

We consider the problem of learning causal directed acyclic graphs from an observational joint distribution. One can use these graphs to predict the outcome of interventional experiments, from which data are often not available. We show that if the observational distribution follows a structural equation model with an additive noise structure, the directed acyclic graph becomes identifiable from the distribution under mild conditions. This constitutes an interesting alternative to traditional methods that assume faithfulness and identify only the Markov equivalence class of the graph, thus leaving some edges undirected. We provide practical algorithms for finitely many samples, RESIT (regression with subsequent independence test) and two methods based on an independence score. We prove that RESIT is correct in the population setting and provide an empirical evaluation.

Original language	Undefined/Unknown
Journal	Journal of Machine Learning Research
Volume	15
Pages (from-to)	2009-2053
Number of pages	45
ISSN	1533-7928
Publication status	Published - Jun 2014
Externally published	Yes

Cite this

@article{ec293cb2d5a1408a9063d191e564f3b9,

title = "Causal Discovery with Continuous Additive Noise Models",

abstract = "We consider the problem of learning causal directed acyclic graphs from an observational joint distribution. One can use these graphs to predict the outcome of interventional experiments, from which data are often not available. We show that if the observational distribution follows a structural equation model with an additive noise structure, the directed acyclic graph becomes identifiable from the distribution under mild conditions. This constitutes an interesting alternative to traditional methods that assume faithfulness and identify only the Markov equivalence class of the graph, thus leaving some edges undirected. We provide practical algorithms for finitely many samples, RESIT (regression with subsequent independence test) and two methods based on an independence score. We prove that RESIT is correct in the population setting and provide an empirical evaluation.",

author = "J. Peters and J.M. Mooij and D. Janzing and B. Sch{\"o}lkopf",

year = "2014",

month = jun,

language = "Udefineret/Ukendt",

volume = "15",

pages = "2009--2053",

journal = "Journal of Machine Learning Research",

issn = "1533-7928",

publisher = "MIT Press",

}

TY - JOUR

T1 - Causal Discovery with Continuous Additive Noise Models

AU - Peters, J.

AU - Mooij, J.M.

AU - Janzing, D.

AU - Schölkopf, B.

PY - 2014/6

Y1 - 2014/6

N2 - We consider the problem of learning causal directed acyclic graphs from an observational joint distribution. One can use these graphs to predict the outcome of interventional experiments, from which data are often not available. We show that if the observational distribution follows a structural equation model with an additive noise structure, the directed acyclic graph becomes identifiable from the distribution under mild conditions. This constitutes an interesting alternative to traditional methods that assume faithfulness and identify only the Markov equivalence class of the graph, thus leaving some edges undirected. We provide practical algorithms for finitely many samples, RESIT (regression with subsequent independence test) and two methods based on an independence score. We prove that RESIT is correct in the population setting and provide an empirical evaluation.

AB - We consider the problem of learning causal directed acyclic graphs from an observational joint distribution. One can use these graphs to predict the outcome of interventional experiments, from which data are often not available. We show that if the observational distribution follows a structural equation model with an additive noise structure, the directed acyclic graph becomes identifiable from the distribution under mild conditions. This constitutes an interesting alternative to traditional methods that assume faithfulness and identify only the Markov equivalence class of the graph, thus leaving some edges undirected. We provide practical algorithms for finitely many samples, RESIT (regression with subsequent independence test) and two methods based on an independence score. We prove that RESIT is correct in the population setting and provide an empirical evaluation.

M3 - Tidsskriftartikel

SN - 1533-7928

VL - 15

SP - 2009

EP - 2053

JO - Journal of Machine Learning Research

JF - Journal of Machine Learning Research

ER -