Extend specialisation to non-Markovian causal settings

Extend the specialisation framework and RadCF counterfactual image-generation mechanism to non-Markovian causal settings in which the exogenous variables may be dependent or unobserved confounding may be present.

Background

The proposed specialisation framework constructs image mechanisms under a causal Markovian assumption. In this setting, the exogenous variables are treated as mutually independent, and the method controls only the observed parent variables supplied to the image mechanism.

The paper notes that clinical data may violate this assumption because of unobserved confounding. Although the appendix probes whether the abducted latent retains parent information, the current method does not model non-Markovian dependencies. Developing an extension for such settings is therefore left unresolved.