Predictive AI could play a vital role in preventing deadly crowd crush incidents like the one in Itaewon, South Korea, by going beyond basic crowd counting to analyze movement patterns and real-time flow dynamics. A research team from KAIST, led by Professor Jae-Gil Lee from the School of Computing, has developed a new AI system capable of accurately forecasting crowd density and potential congestion hotspots. The technology was presented at the Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Unlike traditional approaches that focus only on the number of people in a given area or the direction of movement, the KAIST team’s method uses a time-varying graph model to simultaneously analyze two critical factors: node information (population levels in specific locations) and edge information (the flow of people between areas). This dual analysis is essential because crowd risk isn’t just about how many people are present—it’s about how they are moving. For example, a seemingly low-density area could quickly become dangerous if a large number of people are flowing into it from adjacent zones. By tracking both the current population and the incoming flow, the AI can detect early warning signs of dangerous congestion before it occurs. To achieve this, the team developed a bi-modal learning framework that integrates population counts, movement flows, spatial connections between areas, and temporal changes over time. A key innovation is the use of 3D contrastive learning, which enables the AI to understand not only the spatial layout of an area but also how crowd patterns evolve over time. This allows the system to anticipate future congestion rather than just react to current conditions. The researchers compiled and publicly released six real-world datasets from sources including subway systems in Seoul, Busan, and Daegu, New York City transit data, and COVID-19 case records from South Korea and New York. These datasets were used to train and test the model, which demonstrated up to a 76.1% improvement in prediction accuracy compared to existing state-of-the-art methods. Professor Lee emphasized the technology’s broad social impact, stating it could enhance safety during large public events, improve urban traffic management, and even help monitor and respond to infectious disease spread by tracking movement patterns. He hopes the system will become a valuable tool for protecting public safety in everyday life. The study, titled “Bi-Modal Learning for Networked Time Series,” was published in the 2025 ACM SIGKDD proceedings with DOI: 10.1145/3711896.3736856.