Two Optimal Strategies for Active Learning of Causal Models from Interventional Data

From observational data alone, a causal DAG is only identifiable up to Markov equivalence. Interventional data generally improves identifiability; however, the gain of an intervention strongly depends on the intervention target, that is, the intervened variables. We present active learning (that is, optimal experimental design) strategies calculating optimal interventions for two different learning goals. The first one is a greedy approach using single-vertex interventions that maximizes the number of edges that can be oriented after each intervention. The second one yields in polynomial time a minimum set of targets of arbitrary size that guarantees full identifiability. This second approach proves a conjecture of Eberhardt (2008) indicating the number of unbounded intervention targets which is sufficient and in the worst case necessary for full identifiability. In a simulation study, we compare our two active learning approaches to random interventions and an existing approach, and analyze the influence of estimation errors on the overall performance of active learning.

💡 Research Summary

The paper addresses the fundamental limitation that, from observational data alone, a causal directed acyclic graph (DAG) can only be identified up to its Markov equivalence class. To break this ambiguity, interventional data are required, but the benefit of an intervention depends critically on which variables are intervened upon. The authors formulate this as an active learning (optimal experimental design) problem and propose two distinct strategies tailored to different learning objectives.
The first strategy is a greedy, single‑vertex intervention scheme. At each step the algorithm evaluates, for every candidate variable, the expected number of edge orientations that would become identifiable if that variable were intervened upon. The variable with the highest expected gain is selected, an intervention is performed, and the essential graph is updated. This process repeats until all edges are oriented. The method is computationally efficient—its cost scales linearly with the number of unresolved edges and the number of variables—and it maximizes the immediate information gain at each round.
The second strategy relaxes the restriction to single‑vertex interventions and seeks a minimum‑size set of intervention targets of arbitrary cardinality that guarantees full identifiability of the DAG. By exploiting the structure of the essential graph, the algorithm isolates undirected chain components and ensures that each component receives at least one intervention target. The authors prove that the resulting set is optimal in the worst‑case sense: it matches the conjectured bound by Eberhardt (2008) that ⌈log₂(p)⌉ unbounded intervention targets are both sufficient and, in the worst case, necessary to achieve complete identifiability for a p‑variable system. Importantly, the algorithm runs in polynomial time, making it practical for moderate‑size problems.
A comprehensive simulation study compares the two proposed methods against random interventions and a previously published optimal‑design approach. Performance metrics include (i) the number of newly oriented edges per intervention, (ii) the total number of interventions required for full identifiability, and (iii) robustness to estimation errors introduced during the structure‑learning phase (e.g., false positives/negatives in the essential graph). Results show that the greedy single‑vertex method rapidly reduces uncertainty in early rounds and is relatively robust to noisy estimates, though it typically requires more interventions overall. In contrast, the minimum‑size arbitrary‑target method achieves the smallest possible intervention budget for complete identification but is more sensitive to errors: a single mistaken edge orientation can propagate and increase the required number of interventions.
The paper’s contributions are threefold: (1) it formalizes an information‑theoretic criterion for selecting single‑vertex interventions in a greedy, provably optimal‑per‑step manner; (2) it provides a polynomial‑time algorithm that constructs a provably minimal set of arbitrary‑size interventions, thereby confirming Eberhardt’s conjecture; and (3) it offers an extensive empirical evaluation that quantifies the trade‑offs between intervention cost, speed of learning, and robustness to estimation noise. These insights are directly applicable to experimental sciences where each intervention (e.g., gene knockout, drug treatment) incurs substantial cost, and they guide practitioners in choosing between fast, robust, but potentially more expensive designs and theoretically optimal, cost‑minimal designs that demand higher confidence in intermediate estimates.