Arboreal Neural Network

ICLR 2026 Conference Submission7785 Authors

Keywords: Tabular data analysis, Decision trees, Neural networks, Neural-tree models, Credit-risk dataset, Interpretability

TL;DR: A novel model for tabular data-based predictive modeling that combines the benefits of end-to-end learnability with the interpretability of decision trees.

Abstract: Recent advancements in deep learning and Large Language Models (LLMs) have significantly influenced fields such as Natural Language Processing (NLP), Computer Vision (CV), and audio analysis. However, in the domain of structured tabular data, tree-based models like Gradient Boosted Decision Trees (GBDTs) remain dominant. In this paper, we propose Arboreal Neural Network (ArbNN), a novel model for tabular data that combines the end-to-end learnability of neural networks with the inductive bias and interpretability of decision trees. We achieve this by recasting the discrete logic of a decision tree into a fully differentiable neural structure (dubbed ArborCell), which serves as the basic unit of ArbNN. Moreover, we theoretically prove that the computations in ArborCell are equivalent to a specific self-attention mechanism, providing a new perspective on tree-based models. In addition, we present TabCredit, a large-scale industrial credit-risk dataset designed to enable realistic evaluation under conditions of temporal drift and extensive feature engineering. These two factors are typically lacking in existing benchmarks yet critical for tabular data tasks in real-world scenarios. Empirical evaluations on both public benchmarks and the TabCredit dataset demonstrate ArbNN’s superior performance compared to State-of-the-Art (SOTA) neural network architectures and traditional GBDT methods. ArbNN has been successfully deployed in real-world credit risk management systems, processing millions of loan applications and supporting monthly transactions in billions of U.S. dollars, demonstrating its robustness and industrial value. The code and TabCredit dataset will be released upon acceptance of this paper.

Primary Area: interpretability and explainable AI

Submission Number: 7785