CNN-Based Approaches for Various Types of Tabular Data

Deep learning (DL) includes various architectures, such as deep neural networks (DNNs) and convolutional neural networks (CNNs). DL is very powerful and flexible for non-tabular (non-structured) data (e.g. image, text). However, in tabular data, standard DNNs often do not outperform traditional machine learning (ML) methods such as tree-based models (e.g. random forest, XGBoost). CNNs carry out dimensionality reduction for non-tabular (especially image) data, but may be useful in tabular data too. In this paper, we present a unified framework of one-dimensional CNN (1D-CNN)-based approaches for various types of tabular data, which provides an end-to-end learning framework. We also propose two novel 1D-CNN-based models, i.e. a negative binomial CNN (NB-CNN) model for over-dispersed count data and a Cox-based CNN Self-Attention model for high-dimensional survival data. The predictive performance of the proposed method is evaluated by comparing it with existing ML/DL methods using four types of real tabular data, i.e. a binary response data with high dimensional features, over-dispersed count data, high-dimension survival data, and time-series data with substantial variability. The experimental results show that the proposed methods overall outperform existing ML/DL models. In particular, the NB-CNN achieves lower root mean squared error (RMSE) and higher coefficient of determination (R²) on over-dispersed count data than tree-based methods. Similarly, the Cox-based CNN Self-Attention model yields higher C-index values for high-dimensional survival tasks relative to state-of-the-art approaches.