Interpretable CNN–KAN hybrid architectures for tabular data with synthetic image encoding

Mondragon-Ruiz G.; Liu J.; Castillo-Cara M.; García-Castro R. · Information Processing and Management (2026) · DOI: 10.1016/j.ipm.2026.104954

Resumen

Deep neural networks excel in vision and language but often struggle with tabular data due to weak inductive biases and limited interpretability. While converting tabular data into synthetic images enables Convolutional Neural Networks (CNNs) to exploit spatial feature interactions, this approach typically exacerbates the models’ black-box nature. To achieve both high predictive power and transparent decision-making, we propose an interpretable Hybrid Neural Network that integrates visual reasoning (CNNs applied to synthetic images) with symbolic learning (Kolmogorov-Arnold Networks (KAN) applied to tabular data). Extensive evaluations across diverse datasets demonstrate that our CNN–KAN hybrids achieve highly competitive, and often superior, predictive performance. To audit their internal reasoning, we introduce the Global Feature Score (GFS), a unified metric that aggregates spatial saliency from Grad-CAM with symbolic attributions from the KAN branch, producing consistent feature relevance rankings. By explicitly diagnosing modality dominance, layout artifact sensitivity, and inter-branch agreement, the proposed framework significantly reduces reliance on fragile post-hoc methods. Ultimately, embedding interpretability directly into the architecture provides auditable, cross-verified explanations, establishing a robust foundation for Responsible AI-oriented tabular learning.

Instituciones: Universidad de Lima