Research map / Probabilistic and scientific modeling / Causal discovery
Identifiable Exchangeable Mechanisms for Causal Structure and Representation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the Identifiable Exchangeable Mechanisms (IEM) framework that unifies causal discovery, ICA and causal representation learning under exchangeable non-i.i.d. data, introducing cause and mechanism variability for bivariate causal discovery and a dual identifiability result for TCL.
Problem
Structure (causal discovery) and representation identifiability methods have developed separately; it is unclear what enables both, and identifiability often relies on multi-environment non-i.i.d. data.
Approach
Unify CD, ICA and CRL as a single probabilistic model over exchangeable sequences; relax the causal de Finetti condition so that only cause or only mechanism variability is needed for bivariate causal direction identification, and show a duality between cause and mechanism variability that yields new TCL identifiability under mechanism variability.
Key results
- Shows cause or mechanism variability alone (rather than both) suffices for bivariate causal direction identification, generalizing the causal de Finetti theorem.
- Proves TCL identifiable under mechanism variability via a duality with cause variability.
- Shows joint identification of sources, causal variables and DAG with (dim Z + 1) interventions instead of (2 dim Z + 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Allow cause variability while keeping the effect-given-cause mechanism fixed | Bivariate causal discovery · Exchangeable non-i.i.d. bivariate data | causal structure identifiability | Helped |
| Allow mechanism variability while keeping the cause distribution fixed | Bivariate causal discovery · Exchangeable non-i.i.d. bivariate data | causal structure identifiability | Helped |
| Use exchangeable non-i.i.d. sources with a fixed mixing function | TCL · Auxiliary-variable ICA | source identifiability | Helped |
| Use stationary i.i.d. sources with a stochastic, environment-varying mixing function | TCL · Mechanism-variability formulation | source identifiability | Helped |
| Jointly identify exogenous sources, causal variables, and the DAG using exchangeable data | Causal representation learning · Single-node soft interventions | number of interventions for simultaneous identifiability: (dim Z+1) vs (2 dim Z+1) | Helped |
Surprising findings
- Causal direction can be identified when only one of the cause or mechanism distributions varies, not both.
Limitations
- Bivariate causal discovery result; multivariate extension expected but not shown
- Identifiability results rely on infinite samples; sample complexity not analyzed
- Interventional discrepancy condition holds almost everywhere only under extra restrictions for discrete variables
Experimental setup
Builds on: Causal de Finetti theorem time-contrastive learning identifiable latent neural causal models CauCA
Closest papers by meaning
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.91)
- Identifiable Latent Polynomial Causal Models through the Lens of Change (ICLR 2024 · similarity 0.90)
- Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.89)
- Identifying General Mechanism Shifts in Linear Causal Representations (NeurIPS 2024 · similarity 0.89)
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.88)
- Causal learning with the invariance principle (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.