Linear causal discovery works without common assumptions using multiple data sets
Causal Discovery for Linear Models with Latent Confounding via Multi-domain Identifiable Representation Learning
Machine LearningArtificial Intelligence
Summary
Finding cause-and-effect relationships in data is hard when hidden factors influence things. This paper focuses on figuring out these relationships in simple linear cases even when hidden factors are involved. The authors propose a new method, LinADMG, that uses data from multiple related settings to identify hidden noise and recover causal connections without needing some usual extra assumptions. They test their approach using simulations and partially real data, showing it can find causal structures more completely than earlier methods.
What this means in practice
- •For data scientists: Identify causal relationships in datasets with hidden confounders by using multiple related data sources instead of relying on uncommon data assumptions.
- •For business analysts: Improve understanding of complex system influences in organizational data where hidden factors exist, using causal insights from multi-scenario data.
Authors
Danru Xu, Aapo Hyvärinen, Sara Magliacane
Abstract
Causal discovery with latent confounding is particularly challenging, but a common setting in real-world data. In this work, we focus on linear causal models with latent confounding, represented by acyclic directed mixed graphs (ADMGs). Existing methods that fully identify ADMGs typically assume non-Gaussianity or graphical constraints, such as bow-freeness. In contrast, we propose Linear ADMG Learning (LinADMG), a framework for learning ADMGs for linear causal models through identifiable representation learning. Rather than relying on non-Gaussianity or bow-freeness, LinADMG leverages multi-domain settings to identify the noise variables. We show that the identifiability of the noise variables is sufficient for recovering all direct causal relations, and that, under an additional latent-confounding no-cancellation condition, LinADMG can identify the whole ADMG. We implement LinADMG with four identifiable representation learning approaches from the literature and validate the methods on simulated and semisynthetic datasets.