LG AI Research: A New Approach to Learning Tabular Data

Unlike images or text, Tabular Data is an area where deep learning models struggle to learn complex and irregular functions. Existing deep learning models tend to converge toward smooth functions, making them vulnerable when handling tabular data, which has discrete characteristics.

LG AI Research's Data Intelligence (DI) Lab presented research at NeurIPS 2023 that uses Binning as a Pretext Task to address this issue. This method utilizes an algorithm that converts continuous variables into discrete variables, and is designed to predict Binning class values rather than the input values themselves, through an Autoencoder structure.

The key advantages of the Binning approach are as follows:

• Provides an Inductive Bias that helps learn irregular functions by converting continuous variables into discrete ones
• Effectively integrates the characteristics of tabular data where numerical and categorical variables are mixed
• Increases the efficiency of learning data patterns by grouping similar values into the same Bin
• Secures robustness by allowing unnecessary patterns or minor errors to be ignored

Experimental results showed that this method delivered excellent performance across 25 datasets in binary classification, multi-class classification, and regression, proving comparable or superior performance even when compared to supervised learning-based Tree-based algorithms or state-of-the-art deep learning models.