SEOUL, September 17 (AJP) - The Korea Advanced Institute of Science & Technology (KAIST) has developed an artificial intelligence system that can predict when and where dangerous crowd crushes may occur, raising hopes for preventing disasters such as the Itaewon tragedy that occurred in 2022.
The research, announced on September 17, was led by Professor Lee Jae-gil of KAIST's School of Computing. The study also involved contributions from Professor Yoon Soo-sik of Korea University and Professor Song Hwan-joon of KAIST's Department of Industrial and Systems Engineering. Doctoral students Nam Young-eun and Na Ji-hye of KAIST were among the lead authors. The team presented their findings in August at KDD 2025, one of the world’s top conferences in artificial intelligence and data science.
Professor Lee's team sought to create an AI that could catch the early signs of such crowd disasters. Existing methods often focus only on one factor, such as how many people are currently in a space or how they are moving between areas. The new model combines both.
The researchers introduced a technique called bi-modal learning, which analyzes node data (how many people are in a given location) and edge data (how people are moving between locations) at the same time. To improve accuracy, they added three-dimensional contrastive learning, which allows the AI to capture both spatial relationships and how those relationships shift over time.
This approach makes it possible to spot situations that single-factor models might miss. For example, if one alley is not yet overcrowded but nearby areas are steadily sending more people into it, the AI can flag the location as high-risk before it becomes critical.
To test their system, the researchers built six real-world datasets. These included subway ridership in Seoul, Busan, and Daegu, traffic flow in New York, and COVID-19 case records from South Korea and New York. Across all tests, their model outperformed 16 leading prediction methods. In some cases, it achieved a maximum of 76.1 percent greater accuracy, measured by mean squared error, a standard benchmark for predictive reliability.
"This research shows the importance of developing technology that can have a real social impact," said Professor Lee. "We hope it will contribute to managing large crowds at festivals and events, reducing urban traffic congestion, and even slowing the spread of infectious diseases." The study was presented at the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2025).
reporter
Park Sae-jin
swatchsjp@ajunews.com
Copyright ⓒ Aju Press All rights reserved.
