HyperSDT: HyperNetwork Slide Decision Tree for Interpretable Tabular Learning

Abstract

Recently, substantial progress has been achieved in leveraging deep learning models for tabular data learning. However, despite significant advancements, the predominant focus of these endeavors has been on augmenting the performance of contemporary deep learning models. Consequently, the interpretability of such models is frequently overlooked or rendered secondary, thereby posing a challenge in comprehending their underlying decision-making processes. In this work, we propose a novel HyperNetwork Slide Decision Tree (HyperSDT) approach to achieve interpretable deep learning for tabular data while maintaining a comparable accuracy to state-of-the-art methods. HyperSDT provides a comprehensive interpretable framework with interpretability by using Silde Decision Tree and Decision Transformer together. Our experimental results demonstrate that our framework is competitive with prior baselines under various tabular learning benchmarks while providing better interpretability. The code can be achieved via https://github.com/hunan-create/HyperSDT.