Causal Learning Paradigms: From Prior Causal Structure to Causal Discovery

Introduction Causality is gaining broader attention in the machine learning community, and this report discusses the main paradigms that combine causality with machine learning.

What is Causal Learning? Traditional machine learning focuses on learning correlations from data, which works well for perception tasks but struggles with decision‑making tasks. Causal learning goes beyond correlations by incorporating knowledge about underlying causal mechanisms, leading to better generalization and interpretability.

Why Causal Learning? It offers stronger generalization by eliminating spurious correlations (e.g., camel‑desert background) and provides clearer explanations because causal models align with human reasoning.

Two Main Paradigms

1. Learning with Prior Causal Structure – The causal graph is known beforehand. The key challenge is how to integrate this prior knowledge with existing deep learning models, typically by combining causal structure with neural networks.

2. Learning via Causal Discovery – The causal graph is unknown and must be inferred from data. The challenge lies in obtaining reliable causal knowledge and leveraging it together with deep learning.

Examples of Prior‑Structure Methods – Domain adaptation under conditional shift, where domain information (D) and label (Y) jointly determine features (X). By decoupling D and Y using a causal graph, models can achieve domain‑invariant representations. Network designs may include separate branches for domain and label information, supervised signals for disentanglement, and fine‑tuning of architectures based on corrected causal graphs.

Applications include stable learning, recommendation systems, and computer vision, but they face issues such as verifying the correctness and identifiability of the provided causal graph.

Examples of Causal‑Discovery Methods – Time‑series transfer learning for HVAC control across cities. An LSTM + attention module learns a sparse correlation matrix representing potential causal relations; alignment techniques (e.g., MMD) transfer the learned causal structure to a target domain. Hidden‑variable models treat the causal graph as a latent variable, reconstructing it from observed sequences.

Empirical results show improved performance on tasks like human skeleton behavior transfer and air‑quality prediction, confirming that models with learned causal mechanisms converge faster and generalize better.

Challenges and Future Directions – Causal discovery algorithms often rely on strong assumptions that clash with the open‑world nature of machine learning. Integrating causal discovery with deep learning remains difficult due to mismatched toolsets (independence tests vs. optimization). Future work may explore causal reinforcement learning and stronger causal‑guided interventions to approach general AI.

Conclusion Prior‑structure causal learning enhances generalization and addresses data bias, while causal‑discovery learning improves interpretability and adaptability. Both paradigms still need advances in identifiability, scalability, and seamless integration with deep models.