Causal Discovery under Changing Mechanisms: A Unified Graphical Approach

ICLR 2026 Conference Submission19040 Authors

Keywords: causality, causal discovery

TL;DR: we devise a dynamic causal discovery method under changing mechanisms in a unified causal model

Abstract: A common assumption of general causal mechanisms is invariant across all situations in the environment, just as Newton’s laws of motion are always-valid mechanisms. However, in reality, the causal mechanisms are often partially activated from a partially observed mechanism under specific situations, such as the power mechanism of a hybrid vehicle changes according to the type of energy source available. This brings three following problems: (i) definitely, how to describe these changing mechanisms in a unified causal model, (ii) theoretically, what conditions make the dynamic causal model identifiable, and (iii) methodology, how to learn the model. In response to them, we novelly extend the definition of the directed acyclic graph to the dynamic causal graph with condition labels on edges. We provide the identification when the changing mechanisms follow a linear latent Gaussian dynamic causal model (DynaCM). Building upon these, we devise a five-step algorithm to recover causal mechanisms and reduce condition labels of edges, thereby identifying the dynamic causal graph. Experiments on both synthetic and real-world data demonstrate the effectiveness of our method.

Primary Area: causal reasoning

Submission Number: 19040