CausalKANs: interpretable treatment effect estimation with Kolmogorov-Arnold networks

TL;DR

The paper introduces causalKANs, a framework that converts neural CATE estimators into interpretable, closed-form expressions by substituting backbones with Kolmogorov–Arnoldd Networks and applying pruning and symbolic simplification. It presents a model-agnostic pipeline with explicit budgets to control accuracy loss while yielding auditable formulas for mu0, mu1, and tau, plus interpretable plots. Empirical results show causalKANs often match or exceed neural baselines on standard benchmarks and offer a favorable accuracy–interpretability trade-off, particularly with shallow KAAM variants. The work positions causalKANs as a practical, auditable tool for high-stakes policy, medicine, and economics, with release of code for reproducibility and further adoption.

Abstract

Deep neural networks achieve state-of-the-art performance in estimating heterogeneous treatment effects, but their opacity limits trust and adoption in sensitive domains such as medicine, economics, and public policy. Building on well-established and high-performing causal neural architectures, we propose causalKANs, a framework that transforms neural estimators of conditional average treatment effects (CATEs) into Kolmogorov--Arnold Networks (KANs). By incorporating pruning and symbolic simplification, causalKANs yields interpretable closed-form formulas while preserving predictive accuracy. Experiments on benchmark datasets demonstrate that causalKANs perform on par with neural baselines in CATE error metrics, and that even simple KAN variants achieve competitive performance, offering a favorable accuracy--interpretability trade-off. By combining reliability with analytic accessibility, causalKANs provide auditable estimators supported by closed-form expressions and interpretable plots, enabling trustworthy individualized decision-making in high-stakes settings. We release the code for reproducibility at https://github.com/aalmodovares/causalkans .

Figures (19)

• Figure 1: KAN-ification.
• Figure 2: Overall, causalKANs achieves similar PEHE (lower is better) than causalNNs in IHDP A. * means statistical difference $p<0.05$ in a Wilcoxon paired test. Red brace compares the best causalKAN with the best causalNN, without indicating statistical difference.
• Figure 3: Architectures used for potential outcome regression. Boxes denote layers or backbones (neural networks or KANs); dotted arrows indicate optional hidden layers.
• Figure 4: Radar plot and PDP for variable contribution to CATE, using T-KAAM in ACIC-7.
• Figure 5: PDP for treatment contribution in $\hat{\mu}\xspace({\mathbf{x}}\xspace, {\mathbf{t}}\xspace)$ estimation, using S-KAAM in IHDP-A.