A learning technology that can effectively respond to the phenomenon that reduces the performance of artificial intelligence models has been developed. It is expected to contribute to increasing the possibility of utilizing artificial intelligence in domestic industries and enhancing its performance.

A research team led by Professors Kim Seong-il and Lim Dong-young from the Department of Industrial Engineering and the Graduate School of Artificial Intelligence at UNIST (President Yong-Hoon Lee) has developed a ‘time series learning technique robust to data drift.’

Time series data refers to data collected continuously at regular intervals according to time order. A large number of data used in various industries such as finance, economy, transportation, agriculture, manufacturing, and healthcare are in the form of time series.

Time series data experiences a phenomenon called 'data drift' as external factors affecting data generation change. Data drift refers to the difference between the data used for training an artificial intelligence model and the data in the actual operating environment.

Professor Kim Seong-il added, “When data drift occurs, the performance of time-series learning artificial intelligence models deteriorates,” and “It is a chronic problem that makes it difficult to utilize time-series data in various industries.”

The research team developed a methodology for designing robust neural network structures based on Neural SDEs (Stochastic Differential Equations) that can effectively address these problems.

Neural SDEs are an extension of Neural ODEs, which are a continuous version of the residual neural network model. The research team presented a theoretical basis for a time series Neural SDEs model design methodology that can maintain robustness even in the data drift phenomenon.

The research team presented three neural SDEs models designed according to the methodology: Langevin-type SDE, Linear Noise SDE, and Geometric SDE. The proposed models showed stability and excellent performance when performing various tasks such as interpolation, prediction, and classification on datasets where data drift occurred.

When data drift occurs, a series of engineering processes to quickly detect it and reconstruct and relearn the data entails a great deal of time and cost. The research team has theoretically and experimentally verified all the technologies necessary to make artificial intelligence robust to data drift from the beginning.

Professor Lim Dong-young said, “Recently, there have been frequent cases of performance degradation of time-series artificial intelligence models due to data drift in dynamic data environments,” and added, “The significance of this study lies in the development of a methodology that enables training artificial intelligence to be robust to drift from the beginning, and in theoretically and experimentally verifying its performance.”

First author Researcher Yong-Kyung Oh said, “Through this study, we developed a neural network structure design methodology to prevent the performance of artificial intelligence from deteriorating due to time series data drift.” He added, “We plan to continuously develop time series data drift monitoring technology and learning data reconstruction technology linked to the developed technology so that various domestic companies can utilize them.”

This study was selected as a spotlight paper in the top 5% of the International Conference on Learning Representations (ICLR), a world-renowned international academic society, and is scheduled to be presented in Vienna, Austria in May. In addition, it was supported by the Korea Health Industry Development Institute Biomedical Global Talent Development Project, the National IT Industry Promotion Agency Artificial Intelligence Graduate School Project, and the Ministry of Science and ICT’s National Research Foundation’s Basic Research and Human-Centered-Carbon Neutral Global Supply Chain Research Center.