Causal Generative Neural Networks

TL;DR

Causal Generative Neural Networks (CGNNs) are a novel approach that learns causal structures from observational data using differentiable generative models, outperforming existing methods in various causal discovery tasks.

Contribution

This paper introduces CGNNs, a new neural network-based method that leverages distributional asymmetries and conditional independencies for causal discovery without confounder assumptions.

Findings

CGNNs outperform state-of-the-art methods in cause-effect inference

Effective in identifying v-structures and multivariate causal relations

Works well on both simulated and real datasets

Abstract

We present Causal Generative Neural Networks (CGNNs) to learn functional causal models from observational data. CGNNs leverage conditional independencies and distributional asymmetries to discover bivariate and multivariate causal structures. CGNNs make no assumption regarding the lack of confounders, and learn a differentiable generative model of the data by using backpropagation. Extensive experiments show their good performances comparatively to the state of the art in observational causal discovery on both simulated and real data, with respect to cause-effect inference, v-structure identification, and multivariate causal discovery.