[2509.08617] Towards Interpretable Deep Neural Networks for Tabular Data

Computer Science > Machine Learning arXiv:2509.08617 (cs) [Submitted on 10 Sep 2025 (v1), last revised 26 Mar 2026 (this version, v2)] Title:Towards Interpretable Deep Neural Networks for Tabular Data Authors:Khawla Elhadri, Jörg Schlötterer, Christin Seifert View a PDF of the paper titled Towards Interpretable Deep Neural Networks for Tabular Data, by Khawla Elhadri and 2 other authors View PDF HTML (experimental) 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. Comments: Subjects: Machine Learning (cs.LG) Cite as: arXiv:2509.08617 [cs.LG]   (or arXiv:2509.08617v2 [cs.LG] for this version)   https://doi.org/10.48550/arXiv.2509.08617 Focus to learn more arXiv-issued DOI via DataCite Submission history From: Khawla Elhadri [...