Enhancing the Interpretability of Heterogeneous Treatment Effects using Kolmogorov Arnold Network

• In Event: Addressing Challenges in Causality: Free Will, Mediation and Latent Factors, Heterogeneity, and Variable Selection

Abstract

We introduce a novel post-estimation surrogate modeling approach based on the Kolmogorov-Arnold Network (KAN), named the KAN projection, to enhance interpretability of conditional average treatment effects (CATEs) estimated from black-box models. By decomposing CATE functions into compositions of univariate splines, KAN projection provides multifaceted summaries of heterogeneous treatment effects: visualization of nonlinear effects, detection of key moderators, mapping of interactions, and representations into symbolic expressions. Our simulation demonstrates KAN projection outperforms causal forests in identifying true moderators and recovering covariate effects. Applied to the High School Longitudinal Study of 2009 (HSLS:09), KAN projection reveals interpretable insights into how advanced mathematics course-taking differentially impacts math outcome. This framework bridges flexible algorithms and explainability in educational research.

Authors

• Chenguang Pan, Teachers College, Columbia University

• Yuxuan Li, Teachers College, Columbia University

• Youmi Suk, Teachers College, Columbia University