Gyeongsang National University Professor Beomjin Park’s paper accepted to the International Conference on Machine Learning (ICML)

▸A top-tier international conference in machine learning and artificial intelligence… will be held at COEX in Seoul in July

Revolutionizing tabular data analysis performance with TabularBERT (a BERT model for tabular data)

Beomjin Park, a professor in the Department of Information and Statistics, College of Natural Sciences, at Gyeongsang National University (GNU; President Jinhoe Kwon), has had his paper “TabularBERT: Binning-Based Self-Supervised Learning for Tabular Representation” accepted for publication by the International Conference on Machine Learning (ICML), a top-tier international conference in the field of machine learning and artificial intelligence.

ICML, together with the Conference on Neural Information Processing Systems (NeurIPS) and the International Conference on Learning Representations (ICLR), is a premier top-tier international conference representing the field of machine learning and artificial intelligence, and a leading venue for AI research worldwide. It is also classified as an S-tier conference in the AI/ML field on the Korean Institute of Information Scientists and Engineers’ list of excellent conferences and is included in the National Research Foundation of Korea’s BK21 Computer Science list of outstanding international conferences, reflecting its high prestige within Korea’s research evaluation system.

This study proposes a new artificial intelligence learning method to effectively analyze tabular data used across diverse industries, including finance, healthcare, and manufacturing. While conventional deep learning models achieve excellent performance on image and text data, they have limitations on tabular data, failing to fully capture complex relationships among variables and the structural characteristics of the values.

Professor Beomjin Park’s research team proposed a new method that tokenizes numerical data by converting it into intervals and learns interactions among variables within a Transformer architecture. In particular, through self-supervised learning, it effectively learns the intrinsic structure of the data and is designed to learn representations while preserving the order information of values, thereby improving both performance and stability compared to existing methods.

Professor Park Beom-jin said, “This study improves the representation learning approach for tabular data—the format most commonly used in real-world data analysis—and is expected to contribute to performance improvements across various data analysis and prediction tasks.”

The paper accepted this time is scheduled to be presented at the International Conference on Machine Learning (ICML), which will be held at COEX in Seoul in July 2026.

Photo caption: Professor Park Beom-jin, Department of Information and Statistics, Gyeongsang National University

Content inquiries: Professor Beomjin Park, Department of Information and Statistics, Gyeongsang National University, 055-772-1462

This page is provided via AI machine translation. Some inaccuracies may occur. For official purposes, please refer to the original document.