Kolmogorov-Arnold causal generative models

Area of Science:

• Causal inference and machine learning
• Explainable AI (XAI) and interpretable models

Background:

• Deep causal models offer powerful frameworks for causal reasoning but often lack transparency.
• Opaque mechanisms in complex models hinder auditability, especially in high-stakes applications.

Purpose of the Study:

• Introduce the Kolmogorov-Arnold causal generative model for interpretable causal discovery.
• Enable inspection of model mechanisms through symbolic approximations and visualizations.
• Validate model adequacy using observational data alone.

Main Methods:

• Parameterizing structural equations with Kolmogorov-Arnold Networks for mixed-type tabular data.
• Developing a validation pipeline using distributional matching and independence diagnostics.
• Utilizing symbolic approximations and parent-child visualizations for interpretability.

Main Results:

• The model demonstrates competitive performance, statistically indistinguishable from state-of-the-art on synthetic benchmarks (aggregate post-hoc p-value=0.69).
• Achieved a p-value of 0.18 against the best model on the Sachs' additive dataset.
• Successfully applied to a cardiovascular case study, showcasing interpretable mechanism analysis.

Conclusions:

• Expressive causal generative modeling and functional transparency can be achieved simultaneously.
• The Kolmogorov-Arnold model supports trustworthy deployment in tabular decision-making settings.
• Interpretability workflows facilitate auditing causal systems and parent-child effects.