Improving the Use of Deep Neural Networks with Tabular Data by Exploiting Synthetic Images

Overview

This repository contains the resources and code for the study: "Improving the Use of Deep Neural Networks with Tabular Data by Exploiting Synthetic Images." It benchmarks eight tabular-to-image transformation techniques and evaluates their effectiveness in combination with hybrid architectures (CNN+MLP and ViT+MLP) across diverse datasets and machine learning tasks.

Key Features

• Comprehensive Benchmark: Evaluation of eight transformation techniques for tabular data:
  • TINTO
  • REFINED
  • IGTD
  • FeatureWrap
  • SuperTML
  • BarGraph
  • DistanceMatrix
  • Combination
• Hybrid Architectures: Analysis of CNN+MLP and ViT+MLP combinations.
• Diverse Datasets: Includes regression, binary classification, and multiclass classification tasks.
• Metrics: Performance evaluation using RMSE, Accuracy, Precision, Recall, and F1-score.

Methodology

Datasets

The experiments span a variety of datasets, including:

• Regression: Boston Housing, California Housing, MIMO
• Binary Classification: Dengue/Chikungunya, HELOC
• Multiclass Classification: Covertype, GAS

To download MIMO Dataset: https://ieee-dataport.org/open-access/ultra-dense-indoor-mamimo-csi-dataset

And use convert_csi_to_csv_by_antennas.py

Installation

Clone the repository and install the dependencies:

git clone https://github.com/manwestc/TINTOlib-benchmark