HKAN: Hierarchical Kolmogorov-Arnold Networks for Efficient and Interpretable Feature Interaction Modeling

Jincheng Gao, Shuo Sun

Keywords: Feature Interaction Modeling, Kolmogorov-Arnold Networks, Interpretable Machine Learning, Tabular Data, Function Fitting

Abstract:

Learning complex feature interactions is central to modern machine learning, driving breakthrough performance across domains from structured data analytics to predictive modeling in recommender systems and beyond.
However, despite notable progress, this field still faces three substantial challenges:
i) lack of adaptive topology discovery — models cannot automatically learn which features should interact and at what order; ii) the 'black-box' nature of deep neural
networks with poor explainability of the learned interaction patterns; iii) computational inefficiency due to parameter-heavy architectures with limited scalability.
To address these challenges, we propose a hierarchical sparse framework, namely Hierarchical Kolmogorov-Arnold Network (HKAN), for efficient and interpretable feature interaction modeling with three key aspects:
i) factor-quality-guided evolutionary architecture
search (FG-EAS) to automatically discover data-centric optimal feature grouping
strategies;
ii) hierarchical sparse structure with superior parameter efficiency
iii) B-spline-based univariate function visualization and hierarchical factor structures with end-to-end interpretability
from local to global levels.
To test the predictive and symbolic regression ability of HKAN, we conduct experiments across 10 tabular learning and 2 function fitting tasks. HKAN achieves state-of-the-art (SOTA) or highly competitive performance on the vast majority of datasets while utilizing significantly fewer parameters. Notably, on three of these datasets, it reaches state-of-the-art performance with less than 10\% of the parameters used by the baseline models. Moreover, HKAN can serve as a knowledge discovery tool with excellent explainability (e.g., explicit formulas of data patterns) compared to other black-box baselines, which represents a significant step toward building more trustworthy and accountable AI systems.

Primary Area: interpretability and explainable AI

Submission Number: 17182