Trust Your $\nabla$: Gradient-based Intervention Targeting for Causal Discovery

• DOI: 10.48550/arxiv.2211.13715
• Publication Date: 2022-01-01
• Authors: Mateusz Olko; Michał Zając; Aleksandra Nowak; Nino Scherrer; Yashas Annadani; Sebastian Bauer; Łukasz Kuciński; Piotr Miłoś
• Keywords: Identifiability

Inferring causal structure from data is a challenging task of fundamental importance in science. Observational are often insufficient to identify system's uniquely. While conducting interventions (i.e., experiments) can improve the identifiability, such samples usually and expensive obtain. Hence, experimental design approaches for discovery aim minimize number by estimating most informative intervention target. In this work, we propose novel Gradient-based Intervention Targeting method, abbreviated GIT, that 'trusts' gradient estimator gradient-based framework provide signals acquisition function. We extensive experiments simulated real-world datasets demonstrate GIT performs on par with competitive baselines, surpassing them low-data regime.