In the field of machine learning, the absence of spatial structure in tabular data poses significant limitations on the applicability of Convolutional Neural Networks (CNNs). To address this issue, various works have emerged by converting such data into synthetic images, encapsulating feature similarities within a spatial context and thus broadening the application of CNNs to this type of data. This study delves into various techniques for converting tabular data into synthetic images to develop hybrid models that combine different architectures. Specifically, we propose two hybrid neural networks: Hybrid Neural Network (HyNN), which combines a CNN for analysing synthetic images and a Multi-Layer Perceptron (MLP) for tabular data; and Hybrid Vision Transformer (HyViT), which employs a Vision Transformer (ViT) for analysing synthetic images and a MLP for tabular data. Through experimentation, focused on a regression problem using a MIMO indoor localization large-scale dataset, we benchmarked HyViT and various HyNN configurations against classical regression algorithms. Notably, the HyViT model achieves the lowest Root Mean Squared Error (RMSE) outperforming the HyNN counterparts, and the best classical regression model. These findings underscore the potential of synthetic images and hybrid architectural innovations in enhancing deep neural network models in tabular data.

TINTOlib is a state-of-the-art library that wraps the most important techniques for the construction of Synthetic Images from Sorted Data (also known as Tabular Data).

Original dataset: https://ieee-dataport.org/open-access/ultra-dense-indoor-mamimo-csi-dataset

The dataset used is the DIS scenario with 8 antennas.

Synthetic images and results in the following folder: https://upm365-my.sharepoint.com/:f:/g/personal/jiayun_liu_upm_es/Eqhp7Jj3L3pLnK75Jx66nDQB3zBMp319Nqa4cCjrZmSSxw?e=yBxH2H

For classical models, execute:

python Lazyregressor.py

For hybrid models, run the training scripts inside the training scripts folder. For example:

python training_scripts/HyCNN/DIS_8training_CNN_IGTD.py

Note: The images are generated automatically when executing the training scripts.