FoundCause: Causal Discovery with Latent Confounders from Observational Data

2026-06-16 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

FoundCause is a novel amortized causal discovery model designed to identify directed causal structures and latent confounders from observational data. Trained exclusively on synthetic structural causal models, it directly maps datasets to causal graphs in a single forward pass by learning transferable statistical patterns. Its architecture incorporates a permutation-invariant transformer encoder with alternating attention, pairwise statistical features for causal signals, and a factorized decoder with a triangular refinement module for higher-order causal motifs. FoundCause also features a dedicated confounder module using learnable latent tokens and explicitly handles missing data via masked input. It significantly outperforms 11 classical and 4 other amortized causal discovery methods across 15 real-world datasets, demonstrating a +9.6% improvement in F_1, +1.2% in AUROC, and an 18.9% reduction in structural Hamming distance.

Key takeaway

For AI Scientists and Data Scientists working with observational data, FoundCause offers a significant advancement in causal discovery. You should consider integrating amortized methods like FoundCause to accelerate graph inference and improve accuracy, especially when dealing with latent confounders or missing data. This approach provides a robust alternative to traditional non-amortized techniques, delivering superior performance and efficiency for complex real-world datasets.

Key insights

FoundCause uses amortized learning and specific architectural biases to discover causal graphs and latent confounders from observational data.

Principles

• Amortized learning generalizes causal patterns.
• Inductive biases improve causal discovery.
• Explicitly model latent confounders.

Method

FoundCause maps datasets to causal graphs via a single forward pass using a transformer encoder, statistics-conditioned attention, a factorized decoder, and a triangular refinement module.

In practice

• Apply to real-world observational datasets.
• Use for faster causal inference.
• Model hidden common causes.

Topics

• Causal Discovery
• Latent Confounders
• Amortized Learning
• Transformer Models
• Observational Data
• Missing Data Handling

Best for: Research Scientist, AI Scientist, Data Scientist

Related on AIssential

• Beyond Additivity: Causal Discovery in Location-Scale Noise Models with Hidden Variables — Causal discovery in location-scale noise models with hidden variables is identifiable, even when causes affect both mean and variance.
• MARLIN: Multi-Agent Reinforcement Learning for Incremental DAG Discovery — MARLIN efficiently learns causal DAGs incrementally in online settings using a multi-agent reinforcement learning approach.
• Score-Based Causal Discovery of Latent Variable Causal Models — Score-based causal discovery with identifiability guarantees effectively uncovers latent variable causal structures, outperforming constraint-based methods.
• Multi-Level Causal Embeddings — Causal embeddings generalize abstraction, mapping multiple detailed models into a coarser causal model's sub-systems.
• Consistency evaluation of benchmarks used for causal discovery — Benchmark causal graphs often contain mis-aligned knowledge, hindering causal discovery evaluation.
• Transfer learning for causal forest — Transfer learning improves CATE estimation in causal forests by adapting knowledge across data distribution shifts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.