Causality-Aware Predictions in Static Anticausal Machine Learning Tasks.

Area of Science:

• Machine Learning
• Causal Inference
• Statistics

Background:

• Anticausal machine learning tasks involve prediction where outcomes influence inputs.
• Confounding and mediation are common challenges in predictive modeling.
• Existing methods may struggle to disentangle direct and indirect causal effects.

Purpose of the Study:

• To develop a counterfactual approach for training causality-aware predictive models.
• To enable predictions free from confounding influences.
• To isolate direct or indirect causal effects in the presence of mediators.

Main Methods:

• Training supervised learners on counterfactually simulated inputs.
• Leveraging causal relations to retain relevant associations.
• Focusing initially on linear models for analytical tractability.
• Extending to additive models for nonlinearities.

Main Results:

• The approach generates predictions robust to confounding.
• It can differentiate between direct and indirect causal influences.
• Knowledge of the full causal graph is not required, only confounders/mediators.
• The method shows stability against selection bias-induced dataset shifts.

Conclusions:

• The proposed counterfactual method enhances predictive model reliability in anticausal settings.
• It offers a flexible framework for handling confounding and mediation.
• The approach is validated on synthetic and real-world data, demonstrating practical applicability.