Towards Interpretable Deep Neural Networks for Tabular Data

2025-09-10

AI Summary

To address the limited interpretability of deep neural networks (DNNs) on tabular data—which hinders their deployment in high-stakes domains such as finance and healthcare—this paper proposes XNNTab, a novel architecture. XNNTab innovatively integrates a sparse autoencoder with latent-space semantic mapping to learn a univocal, semantically annotatable feature dictionary, enabling linear decomposition and full transparency of the prediction process. Key technical contributions include latent-feature disentanglement, automated semantic annotation, and a linearly interpretable prediction structure. Evaluated on multiple standard tabular benchmarks, XNNTab matches or surpasses the predictive performance of state-of-the-art black-box models (e.g., TabNet, MLP) and traditional methods (e.g., XGBoost), while providing complete, human-understandable, layer-wise explanations. This work represents the first substantive breakthrough toward high-performance, fully transparent, and inherently interpretable DNNs for tabular data.

Technology Category

• Machine Learning: Transparent, Interpretable, Explainable ML
• Humans and AI: Explainable AI (XAI) for Human Understanding
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Tabular data is the foundation of many applications in fields such as finance and healthcare. Although DNNs tailored for tabular data achieve competitive predictive performance, they are blackboxes with little interpretability. We introduce XNNTab, a neural architecture that uses a sparse autoencoder (SAE) to learn a dictionary of monosemantic features within the latent space used for prediction. Using an automated method, we assign human-interpretable semantics to these features. This allows us to represent predictions as linear combinations of semantically meaningful components. Empirical evaluations demonstrate that XNNTab attains performance on par with or exceeding that of state-of-the-art, black-box neural models and classical machine learning approaches while being fully interpretable.

Problem

• Interpretable deep neural networks for tabular data
• Overcoming black-box nature of DNN predictions
• Assigning human-interpretable semantics to learned features

Innovation

• Sparse autoencoder learns monosemantic feature dictionary
• Automated method assigns human-interpretable semantics to features
• Predictions become linear combinations of interpretable components