LG AI Research Presents Tabular Domain Research Results at ICML 2024

LG AI Research's Data Intelligence (DI) Lab presented the paper 'Binning as a Pretext Task: Improving Self-Supervised Learning in Tabular Domains' at ICML 2024. This research proposes a method that overcomes the limitations of Tabular Learning by using a Binning Algorithm to convert numerical variables into categorical variables, and learning representations through a Pretext Task that predicts these converted values.

In this paper, performance was verified not only with existing MLP or FT-Transformer but also using T2G-Former, a state-of-the-art model architecture. It also included various experimental results that visualize the characteristics of tabular data to demonstrate the performance improvement effect.

In addition, the CHARMS methodology, a study on Image-tabular Multimodal Learning that combines image and tabular data, was also introduced. This methodology uses Optimal Transport to maximize the mutual information between image channels and tabular data attributes, solving the alignment problem between different modalities and improving prediction performance.