Research map / Probabilistic and scientific modeling / Causal discovery

Bayesian Active Learning for Bivariate Causal Discovery

ICML 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Interventional causal discovery

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.

Frames bivariate causal direction discovery as a hypothesis test and selects sequential interventions by optimizing a Bayes-factor-based probability of decisive and correct evidence, solved with dynamic programming.

Problem

Active intervention design for determining causal direction between two variables; information-gain objectives may not reliably measure correctness of directional conclusions under a fixed intervention budget.

Approach

Replace information-gain criteria with a Bayes-factor objective (probability of decisive and correct evidence, PDC) for hypothesis testing of X→Y vs Y→X, extended to a discounted multi-step objective whose discrete-case optimization is made polynomial via count-vector dynamic programming.

Key results

• In weak-causation binary setting (|p0-p1| in [0.1,0.2]), the method achieves recall of about 0.627 at 1000 observational samples with type I error about 0.04, outperforming baselines in recall.
• On 30-vertex tree graphs with delta=0.05, the method needs about 140 interventions on average, roughly 10% fewer than information-theoretic baselines.
• In the switch-light embodied task with |p0-p1| in (0.2,0.3), the method needs 32.11 interventions versus 33.51 for InfoGain, the best baseline.

Measured effects reported in the paper

Surprising findings

• With continuous additive noise models, all methods perform comparably, since causal direction is often identifiable from observational data alone.
• Random intervention shows competitive average performance in some tree settings, but its distribution is bimodal with high variance.

Limitations

• Extension to multivariate graphs relies on identifiability of the interventional distribution p(y|do(x)), which limits generality.
• Continuous-setting results show little advantage over baselines because ANM direction is often identifiable from observational data.
• Embodied AI experiments are on a single simulated switch-light task.

Experimental setup

Datasets: synthetic binary bivariate SCM data synthetic random tree graphs (30 vertices) TongSim-based switch-light simulation environment

Benchmarks: synthetic bivariate binary causal discovery synthetic tree causal graph learning switch-light reasoning in TongSim

Metrics: type I error rate recall PDC average number of interventions

Baselines: Rand InfoGain Multi-step InfoGain Greenewald et al. (2019)

Builds on: information-gain active causal discovery (Tigas et al., 2022; Toth et al., 2022) Bayes factor model selection (Kass & Raftery, 1995) tree-structured intervention design (Greenewald et al., 2019) multi-step lookahead sequential design (Gonzalez et al., 2016)

Closest papers by meaning

• Bivariate Causal Discovery using Bayesian Model Selection (ICML 2024 · similarity 0.91)
• A Meta-Learning Approach to Bayesian Causal Discovery (ICLR 2025 · similarity 0.89)
• Sample Efficient Bayesian Learning of Causal Graphs from Interventions (NeurIPS 2024 · similarity 0.88)
• Distinguishing Cause from Effect with Causal Velocity Models (ICML 2025 · similarity 0.88)
• CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.88)
• Characterization and Learning of Causal Graphs from Hard Interventions (NeurIPS 2025 · similarity 0.88)