Rethinking the framework constructed by counterfactual functional model.

Area of Science:

• Artificial Intelligence
• Causal Inference
• Machine Learning

Background:

• Causal inference, particularly counterfactual reasoning, is vital for advancing artificial intelligence.
• Existing methods face theoretical challenges in probabilistic evaluation of counterfactual queries and assumption validity.

Purpose of the Study:

• To address limitations in evaluating counterfactual queries within functional models.
• To propose a relaxed assumption for estimating causal effects, overcoming limitations of strong existing assumptions.

Main Methods:

• Analyzed ambiguities in functional models for counterfactual query evaluation when probability solutions are zero.
• Introduced a novel, milder assumption named M-TUA (Mild Treatment-Unit Additivity), inspired by physics (damped vibration equation).

Main Results:

• Identified specific limitations of functional models leading to ambiguous results in counterfactual query evaluation.
• Demonstrated that M-TUA reduces constraint strength compared to traditional assumptions like treatment-unit additivity, offering a more scientifically grounded approach.

Conclusions:

• The proposed M-TUA assumption provides a more flexible and realistic framework for causal effect estimation.
• Resolving these fundamental issues enhances the reliability and applicability of causal inference in AI.