Deep Neural Networks and Tabular Data: A Survey

  • 2021-10-05 09:22:39
  • Vadim Borisov, Tobias Leemann, Kathrin Se├čler, Johannes Haug, Martin Pawelczyk, Gjergji Kasneci
  • 23

Abstract

Heterogeneous tabular data are the most commonly used form of data and areessential for numerous critical and computationally demanding applications. Onhomogeneous data sets, deep neural networks have repeatedly shown excellentperformance and have therefore been widely adopted. However, their applicationto modeling tabular data (inference or generation) remains highly challenging.This work provides an overview of state-of-the-art deep learning methods fortabular data. We start by categorizing them into three groups: datatransformations, specialized architectures, and regularization models. We thenprovide a comprehensive overview of the main approaches in each group. Adiscussion of deep learning approaches for generating tabular data iscomplemented by strategies for explaining deep models on tabular data. Ourprimary contribution is to address the main research streams and existingmethodologies in this area, while highlighting relevant challenges and openresearch questions. To the best of our knowledge, this is the first in-depthlook at deep learning approaches for tabular data. This work can serve as avaluable starting point and guide for researchers and practitioners interestedin deep learning with tabular data.

 

Quick Read (beta)

loading the full paper ...