Scalable Contrastive Causal Discovery under Unknown Soft Interventions

A team of researchers including Mingxuan Zhang, Khushi Desai, Sopho Kevlishvili, and Elham Azizi has introduced a novel AI model for causal discovery, addressing a fundamental limitation in the field. Traditional observational methods can only identify causal relationships up to a Markov equivalence class, leaving ambiguity. While interventions can clarify these relationships, real-world interventions are often 'soft'—affecting multiple unknown targets—and data is frequently limited to just a single intervention regime. This new model, detailed in the paper 'Scalable Contrastive Causal Discovery under Unknown Soft Interventions,' is specifically designed for this challenging but common scenario where the underlying causal structure is shared but interventions are unknown.

The model's technical innovation lies in its scalable approach that aggregates partial directed acyclic graphs (PDAGs) from data subsets and applies contrastive cross-regime orientation rules. This process constructs a globally consistent maximal PDAG, a more complete causal graph, while adhering to Meek's rules of graphical orientation. The authors provide theoretical proofs that their model is sound within a restricted equivalence class and can asymptotically recover the identifiable PDAG. Crucially, experiments on synthetic data demonstrate that this contrastive method can orient a greater number of causal edges compared to non-contrastive techniques, improves overall structure recovery, generalizes to unseen graphs, and scales effectively to larger problems. This represents a significant step toward practical causal AI in messy, real-world environments where experimental control is limited.