PermGrad: Interpretable Hybrid Neural Networks with Synthetic Images for Tabular Data

Deep learning models have achieved remarkable success in vision and language, yet their application to tabular data remains challenging. This work introduces PermGrad, a Hybrid Neural Network (HyNN) framework that integrates tabular and image-based representations to enhance both predictive performa...

Descripción completa

Detalles Bibliográficos
Autor: Escalera González, Felipe
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universidad Nacional de Educación a Distancia
Repositorio:e-spacio. Repositorio Institucional de la UNED
Idioma:inglés
OAI Identifier:oai:e-spacio.uned.es:20.500.14468/31806
Acceso en línea:https://hdl.handle.net/20.500.14468/31806
Access Level:acceso abierto
Palabra clave:1203.17 Informática
Deep Learning
Hybrid Neural Network
Synthetic images
TINTOlib
Tabular-to-images
Interpretability
Descripción
Sumario:Deep learning models have achieved remarkable success in vision and language, yet their application to tabular data remains challenging. This work introduces PermGrad, a Hybrid Neural Network (HyNN) framework that integrates tabular and image-based representations to enhance both predictive performance and interpretability. Tabular data are transformed into synthetic images using the TINTO methodology, enabling convolutional neural networks (CNNs) to capture local feature interactions, while a multilayer perceptron (MLP) branch models global relationships. Interpretability is achieved through three complementary components: permutation-based feature importance for the MLP branch, Grad-CAM saliency mapping for the CNN branch, and a novel PermGrad score, which integrates branch-level contributions via pruning-based analyses. Extensive experiments on five heterogeneous datasets, covering classification and regression tasks, demonstrate that the proposed hybrid approach provides competitive performance against classical machine learning baselines, while offering transparent, featurelevel, and branch-level contributions. The study highlights design trade-offs between interpretability and computational cost, and outlines directions for future extensions including transformerbased backbones and advanced fusion strategies.