Reproduce “Causal Representation Learning Made Identifiable by Grouping of Observational Variables” on OpenResearch

About this paper

Hiroshi Morioka, Aapo Hyvärinen

G-CaRL uses known, non-overlapping observation groups to recover latent variables and inter-group causal links from passive data. It independently shuffles group samples and trains a nonlinear classifier to distinguish aligned from mismatched combinations. Under stated identifiability conditions, representations and some graph structure are recoverable up to ambiguities. Experiments outperform tested baselines, but identify only inter-group edges.

Key insights

• The grouping of observational variables provides sufficient structural information to achieve identifiability for both latent representations and inter-group causal graphs in an unsupervised setting.
• Identifiability is achieved for complex scenarios including nonlinear functions, cyclic causal graphs, and latent confounders, without requiring temporal data or interventions.
• The self-supervised G-CaRL framework is statistically consistent, ensuring its learned representations and causal parameters converge to their true values.

Summary by alphaXiv.