New Method Discovers Unsupervised Causal Abstractions from Data

A new research paper introduces a novel technique for automatically identifying high-level causal relationships within complex systems, based solely on observed data. Unlike traditional methods that require experts to hypothesize and test causal models, this approach learns them directly. The core idea involves using principles from low-rank causal discovery to infer underlying latent variables. These latent variables are then shown to naturally form a 'causal abstraction,' meaning they accurately capture the interventional behavior of the more detailed, lower-level system. The paper provides theoretical identifiability results for these latent variables and proposes a practical objective function to facilitate the learning of these high-level causal models. This marks a significant step towards more autonomous causal discovery.