UvA-DARE ( Digital Academic Repository ) Causal Discovery with Continuous Additive Noise Models
By Jan Peters, Joris M. Mooij, Dominik Janzing, Dominik Janzing, Bernhard Schölkopf. 2014
682 citations recorded by the source corpus.
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.
Subfields: Artificial Intelligence, Statistics and Probability, Causal Inference & Discovery
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Bernhard Schölkopf
- Yoshua Bengio
- Sergey Levine
- Mohit Bansal
- Quoc V. Le
- Percy Liang
- Jianfeng Gao
- Ruslan Salakhutdinov
In Statistics and Probability
- Bernhard Schölkopf
- Jan Peters
- David Lopez-Paz
- Arthur Gretton
- Kun Zhang
- Dominik Janzing
- Peter Buhlmann
- Dino Sejdinovic
In Causal Inference & Discovery
- Bernhard Schölkopf
- David Lopez-Paz
- Jan Peters
- Dominik Janzing
- Kun Zhang
- Joris M. Mooij
- Peter Buhlmann
- Stefan Bauer
Citations in the corpus
- Gradient-Based Neural DAG Learning (2019, methodology): “There are many examples of graph identifiability results for continuous variables [11, 9, 14, 17] as well as for discrete variables [10].”
- Toward Causal Representation Learning (2021, background): “While causal structure discovery is hard if we have only two observables [190], the case of more observables is surprisingly easier, the reason being that, in that case, there are nontrivial conditional independence properties [52],”
- Discovering Causal Signals in Images (2016, background): “In the absence of any assumptions, the determination of causal relations between random variables given samples from their joint distribution is fundamentally impossible [22, 23].”
- Towards Causal Representation Learning (2021, background): “[237, 188, 229, 113, 174, 187, 139, 17, 246, 277, 175, 123, 186, 176, 36, 82, 161], learning causal relations frequently requires collecting data from multiple environments, or the ability to perform interventions [251].”
- Causality for Machine Learning (2019, background): “…X,Y generated by an additive noise model, one cannot fit an additive noise model in the opposite direction (i.e., with the roles of X and Y interchanged) (Hoyer et al., 2009; Mooij et al., 2009; Peters et al., 2014; Kpotufe et al., 2014; Bauer et al., 2016), cf. also the work of Sun et al. (2006).”
- Causal Generative Neural Networks (2017, methodology): “We compare CGNN to the PC algorithm [Spirtes et al. , 2000], the score-based methods GES [Chickering, 2002], LiNGAM [Shimizu et al. , 2006], causal additive model (CAM) [Peters et al. , 2014] and with the pairwise methods ANM and Jarfo.”
- Invariant Causal Prediction for Sequential Data (2017, methodology): “Many procedures in this field only use observational data (e.g., Spirtes, Glymour, and Scheines 2000; Chickering 2002; Shimizu et al. 2006; Janzing et al. 2012; van de Geer and Bühlmann 2013; Peters et al. 2014; Bühlmann, Peters, and Ernest 2014).”
- Kernel‐based tests for joint independence (2016, methodology): “A more efficient method, which is based on a similar idea, is the regression with subsequent independence test algorithm that was described in Peters et al. (2014), section 4.1.”
- Causal inference by using invariant prediction: identification and confidence intervals (2015, background): “, 2006), non-linearities (Hoyer et al., 2009; Peters et al., 2014; Bühlmann et al., 2014), equal error variances (Peters and Bühlmann, 2014) or shared information between the regression function and target variable (Janzing et al.”
- Computing functions of random variables via reproducing kernel Hilbert space representations (2015, positive): “Every graphical model w.r.t. a DAG can be expressed as such a structural equation model with suitable functions and noise terms [e.g., Peters et al. (2014)].”
- Distinguishing Cause from Effect Using Observational Data: Methods and Benchmarks (2014, background): “More details and extensions to the multivariate case can be found in (Hoyer et al., 2009; Peters et al., 2014).”
- CAM: Causal Additive Models, high-dimensional order search and penalized regression (2013, methodology): “As in Theorem 1, we make the assumption ξp > 0 in (9): its justification, however, is somewhat less backed up because the identifiability results from [26] and Lemma 3 do not carry over immediately.”
- Causal Discovery with Reinforcement Learning (2019, methodology): “, we can fit each variable x iagainst all the rest variables, and use only the RSS iterms but ignore the additive penalty on the number of edges. With the independence based score function proposed by Peters et al. (2014), we may simply set S L= 0. The minimum term h min, as previously mentioned, may”
- Optimal experimental design via Bayesian optimization: active causal structure learning for Gaussian process networks (2019, methodology): “well as the number of DAGs as mentioned before). To improve tractability, a combination of our method with constraint-based approaches [6], and regression with subsequent independence-testing (RESIT) [38] seems to be an interesting direction for future work. Applicability and suggested use cases Due”
- Causal inference in partially linear structural equation models (2016, methodology): “The authors provide an example where two DAGs are distribution equivalent if one of the nonlinear functions is replaced by a linear function [18], Example 26.”
- Joint Causal Inference from Multiple Contexts (2016, background): “, 2014), algebraic constraints in the linear-Gaussian case (van Ommen and Mooij, 2017), non-Gaussianity in linear models (Kano and Shimizu, 2003), and non-additivity of noise in nonlinear models (Peters et al., 2014) can also be exploited.”