PermGrad: Interpretable Hybrid Neural Networks with synthetic images for tabular data

Escalera-González F.; Castillo-Cara M.; Rincón-Zamorano M.; Orozco-Barbosa L. · Knowledge-Based Systems (2026) · DOI: 10.1016/j.knosys.2026.116507

Resumen

Deep learning has achieved notable progress in vision and language domains, yet its application to tabular data remains constrained by the lack of inherent spatial structure and the profound challenge of extracting transparent explanations. This work introduces a Hybrid Neural Network (HyNN) framework focused on robust interpretability, combining symbolic reasoning on original tabular features with spatial representations derived from synthetic images. Tabular data are transformed into TINTO-generated synthetic images, enabling a CNN branch to exploit local feature interactions while an MLP branch captures global dependencies. To decode this hybrid decision process, we propose PermGrad: a unified attribution pipeline integrating permutation-based importance (MLP), Grad-CAM saliency (CNN), and a macroscopic pruning-based weighting scheme that scales feature relevance by each branch's empirical contribution. Experiments across five heterogeneous datasets reveal that the HyNN achieves predictive performance highly competitive with SOTA deep tabular architectures (e.g., FT-Transformer). Crucially, PermGrad exposes the internal behavioural dynamics of the hybrid model: the CNN leverages the spatial layout to dominate high-dimensional contexts, while the MLP acts as a vital ”symbolic anchor” during training. Even when its explicit inference-time attribution is negligible, the auxiliary MLP stabilizes gradient flow, prevents optimization collapse, and yields consistently lower test losses. Overall, PermGrad provides a unified, scalable, and class-aware view of feature importance, demonstrating that the strategic use of synthetic images in hybrid architectures can bridge the gap between competitive predictive power and transparent, modality-aware explainability.

Instituciones: Universidad de Lima