Theory-based causal induction.

Area of Science:

• Cognitive Science
• Machine Learning
• Statistics

Background:

• Causal induction from observations is a fundamental challenge in science and human learning.
• People effectively infer causal structures from diverse data types, yet these processes are often studied in isolation.
• Identifying unobservable causal mechanisms from sparse data remains a significant inductive problem.

Purpose of the Study:

• To present a unified computational framework for causal induction.
• To model diverse forms of causal learning within a common language.
• To analyze the role of domain-specific prior knowledge in causal inference.

Main Methods:

• Computational-level analysis of causal induction.
• Development of a framework integrating statistical inference and abstract causal theories.
• Identification of key aspects of prior knowledge constraining causal models.

Main Results:

• Causal induction is framed as domain-general statistical inference guided by domain-specific abstract causal theories.
• A common language is established to model various causal learning modes.
• Three key aspects of prior knowledge (ontology, plausibility, functional form) are identified as crucial constraints.

Conclusions:

• Abstract causal theories provide essential constraints for inducing causal models from sparse data.
• The proposed framework unifies diverse causal learning mechanisms.
• Understanding prior knowledge is key to explaining human causal induction capabilities.