Structural Agnostic Modeling: Adversarial Learning of Causal Graphs

By Diviyan Kalainathan, Olivier Goudet, Isabelle M Guyon, David Lopez-Paz, Michèle Sebag. 2018

121 citations recorded by the source corpus.

A new causal discovery method, Structural Agnostic Modeling (SAM), is presented in this paper. Leveraging both conditional independencies and distributional asymmetries in the data, SAM aims at recovering full causal models from continuous observational data along a multivariate non-parametric setting. The approach is based on a game between $d$ players estimating each variable distribution conditionally to the others as a neural net, and an adversary aimed at discriminating the overall joint conditional distribution, and that of the original data. An original learning criterion combining distribution estimation, sparsity and acyclicity constraints is used to enforce the end-to-end optimization of the graph structure and parameters through stochastic gradient descent. Besides the theoretical analysis of the approach in the large sample limit, SAM is extensively experimentally validated on synthetic and real data.

Subfields: Artificial Intelligence, 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

In Causal Inference & Discovery

Citations in the corpus

Gradient-Based Neural DAG Learning (2019, background): “SAM (Kalainathan et al., 2018) learns conditional NN generators using adversarial losses but does not enforce acyclicity.”
Masked Gradient-Based Causal Structure Learning (2019, methodology): “A more related work is SAM [12] that also considered binary structural gates and the smooth acyclicity constraint.”
D’ya Like DAGs? A Survey on Structure Learning and Causal Discovery (2021, methodology): “SAM addresses these two limitations with the use of adversarial training [73], and by making the mechanism which optimizes the DAG part of end-to-end training.”
DiBS: Differentiable Bayesian Structure Learning (2021, methodology): “The mask representation in (E.1) and (E.3) is equivalent to the concept of a structural gate used by Kalainathan et al. [63].”
Differentiable Causal Discovery Under Latent Interventions (2022, background): “The works by Ng et al. (2020), Kalainathan et al. (2020), and Brouillard et al. (2020) treat the adjacency matrix as a random variable and relax the score in Eq.”