Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MetaCaDI: A Meta-Learning Framework for Scalable Causal Discovery with Unknown Interventions

Hans Jarett Ong, Yoichi Chikahara, Tomoharu Iwata

TL;DR

MetaCaDI tackles the challenge of jointly discovering a shared causal graph and unknown intervention targets under few-shot conditions by reframing the problem as Bayesian meta-learning. It introduces an analytical adaptation mechanism that avoids gradient-based bilevel optimization, coupled with a differentiable DAG sampler (DP-DAG) and a differentiable intervention predictor to infer $A$ and $m_t$ under an ANM-based likelihood. The method provides posterior uncertainty over the graph and targets and demonstrates superior performance in both causal-graph recovery and intervention-target prediction on synthetic and SERGIO datasets, with notable speed advantages over gradient-based baselines. These results suggest a practical path toward reliable, real-time causal inference in systems where interventions are costly or partly unknown, enabling faster scientific discovery and fault localization. The work also identifies future directions to broaden model identifiability beyond ANMs and to extend applicability to more complex SCM families.

Abstract

Uncovering the underlying causal mechanisms of complex real-world systems remains a significant challenge, as these systems often entail high data collection costs and involve unknown interventions. We introduce MetaCaDI, the first framework to cast the joint discovery of a causal graph and unknown interventions as a meta-learning problem. MetaCaDI is a Bayesian framework that learns a shared causal graph structure across multiple experiments and is optimized to rapidly adapt to new, few-shot intervention target prediction tasks. A key innovation is our model's analytical adaptation, which uses a closed-form solution to bypass expensive and potentially unstable gradient-based bilevel optimization. Extensive experiments on synthetic and complex gene expression data demonstrate that MetaCaDI significantly outperforms state-of-the-art methods. It excels at both causal graph recovery and identifying intervention targets from as few as 10 data instances, proving its robustness in data-scarce scenarios.

MetaCaDI: A Meta-Learning Framework for Scalable Causal Discovery with Unknown Interventions

TL;DR

MetaCaDI tackles the challenge of jointly discovering a shared causal graph and unknown intervention targets under few-shot conditions by reframing the problem as Bayesian meta-learning. It introduces an analytical adaptation mechanism that avoids gradient-based bilevel optimization, coupled with a differentiable DAG sampler (DP-DAG) and a differentiable intervention predictor to infer and under an ANM-based likelihood. The method provides posterior uncertainty over the graph and targets and demonstrates superior performance in both causal-graph recovery and intervention-target prediction on synthetic and SERGIO datasets, with notable speed advantages over gradient-based baselines. These results suggest a practical path toward reliable, real-time causal inference in systems where interventions are costly or partly unknown, enabling faster scientific discovery and fault localization. The work also identifies future directions to broaden model identifiability beyond ANMs and to extend applicability to more complex SCM families.

Abstract

Uncovering the underlying causal mechanisms of complex real-world systems remains a significant challenge, as these systems often entail high data collection costs and involve unknown interventions. We introduce MetaCaDI, the first framework to cast the joint discovery of a causal graph and unknown interventions as a meta-learning problem. MetaCaDI is a Bayesian framework that learns a shared causal graph structure across multiple experiments and is optimized to rapidly adapt to new, few-shot intervention target prediction tasks. A key innovation is our model's analytical adaptation, which uses a closed-form solution to bypass expensive and potentially unstable gradient-based bilevel optimization. Extensive experiments on synthetic and complex gene expression data demonstrate that MetaCaDI significantly outperforms state-of-the-art methods. It excels at both causal graph recovery and identifying intervention targets from as few as 10 data instances, proving its robustness in data-scarce scenarios.
Paper Structure (29 sections, 13 equations, 2 figures, 1 table)

This paper contains 29 sections, 13 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Causal Discovery from Observational and Interventional Data
  4. Existing Approaches and Motivation
  5. Meta-Learning for Causal Discovery
  6. Problem Setup
  7. Meta-Learning Strategy for Effective Posterior Inference
  8. Proposed Method
  9. Variational Model Components
  10. Differentiable Causal DAG Sampler
  11. Differentiable Intervention Target Predictor
  12. Likelihood Model
  13. Modeling for Efficient Parameter Learning
  14. Feature Design for Intervention Target Prediction
  15. Closed-Form Solvers for Likelihood Fitting
  16. ...and 14 more sections

Figures (2)

  • Figure 1: Our meta-learning setup. Meta-training Phase: Given a collection of datasets $\mathcal{D}_\text{train} = \{D_1, \dots, D_T\}$, MetaCaDI infers shared causal DAG adjacency matrix $\mathbf{A}$ from $\mathcal{D}_\text{train}$, as well as intervention target $\textbf{m}_t$ for each $D_t$. Meta-test Phase: Using a new, small dataset $D'_{t'}$, MetaCaDI performs fine-tuning to infer its specific intervention targets.
  • Figure 2: Performance comparison on the SERGIO and Synthetic datasets, with results averaged over 20 independent simulations per method. Error bars indicate the standard deviation. The y-axis for E-SID is on a logarithmic scale for easier comparison in the presence of outliers. $\downarrow$ and $\uparrow$ indicate "lower is better" and "higher is better", respectively.

Theorems & Definitions (1)

  • Remark 3.1: Identifiability