Causal Learning Paradigms: From Prior Causal Structure to Causal Discovery

The growing interest in causality within machine learning has led to the emergence of causal learning, which extends traditional correlation‑based approaches by incorporating knowledge of underlying causal mechanisms to improve generalization and interpretability.

The presentation is organized into three sections: (1) an overview of what causal learning is, (2) methods that leverage prior causal structures, and (3) methods that rely on causal discovery.

Causal learning differs from conventional machine learning by focusing on the data-generating causal process; it can mitigate spurious correlations (e.g., the camel‑desert background issue) and provide more robust decision‑making in dynamic environments.

Two major paradigms are described. The first assumes a known causal graph and combines it with deep learning (causal‑structure + DL) to address problems such as domain adaptation, conditional shift, and the integration of causal priors into network architectures.

The second paradigm tackles scenarios where the causal graph is unknown. It employs causal discovery techniques to infer latent causal structures, then integrates these structures with deep models (causal‑discovery + causal‑structure + DL). Examples include time‑series transfer for HVAC control, skeleton‑based action recognition, and air‑quality prediction.

Challenges remain: causal discovery often relies on strong assumptions that may not hold in open‑world settings, and reconciling the optimization‑driven nature of deep learning with the statistical‑test‑driven nature of causal discovery is non‑trivial.

In summary, causal learning aims to fuse causal reasoning with AI to achieve better generalization, interpretability, and potential pathways toward general‑purpose AI, though further work is needed on identifiability, supervision requirements, and seamless integration with deep learning frameworks.