New Study Unveils Cutting-Edge Time Series Deep Learning Technique for Optimal Performance in AI Models

Abstract
Irregular sampling intervals and missing values in real-world time series data present challenges for conventional methods that assume consistent intervals and complete data. Neural Ordinary Differential Equations (Neural ODEs) offer an alternative approach, utilizing neural networks combined with ODE solvers to learn continuous latent representations through parameterized vector fields. Neural Stochastic Differential Equations (Neural SDEs) extend Neural ODEs by incorporating a diffusion term, although this addition is not trivial, particularly when addressing irregular intervals and missing values. Consequently, careful design of drift and diffusion functions is crucial for maintaining stability and enhancing performance, while incautious choices can result in adverse properties such as the absence of strong solutions, stochastic destabilization, or unstable Euler discretizations, significantly affecting Neural SDEs’ performance. In this study, we propose three stable classes of Neural SDEs: Langevin-type SDE, Linear Noise SDE, and Geometric SDE. Then, we rigorously demonstrate their robustness in maintaining excellent performance under distribution shift, while effectively preventing overfitting. To assess the effectiveness of our approach, we conduct extensive experiments on four benchmark datasets for interpolation, forecasting, and classification tasks, and analyze the robustness of our methods with 30 public datasets under different missing rates. Our results demonstrate the efficacy of the proposed method in handling real-world irregular time series data.

A team of researchers, led by Professor Sungil Kim and Professor Dongyoung Lim from the Department of Industrial Engineering and the Artificial Intelligence Graduate School at UNIST, has unveiled a groundbreaking time series machine learning technique designed to address data drift challenges. This innovative approach effectively handles irregular sampling intervals and missing values in real-world time series data, offering a robust solution for ensuring optimal performance in artificial intelligence (AI) models.

Time series data, characterized by continuous chronological data collection, is prevalent across various industries, such as finance, transportation, and healthcare. However, the presence of data drifts—changes in external factors influencing data generation—presents a significant hurdle to leveraging time series data effectively in AI models.

Professor Kim emphasized the critical necessity of combatting the detrimental effects of data drift on time series learning models, stressing the urgency of addressing this persistent issue that impedes the optimal utilization of time series data in AI models. In response to this challenge, the research team has introduced a novel methodology leveraging Neural Stochastic Differential Equations (Neural SDEs) to construct resilient neural network structures, capable of mitigating the impacts of data drifts.

Neural SDEs, an extension of neural Ordinary Differential Equations (ODEs), represent continuous versions of residual neural network models and form the cornerstone of the team’s innovative approach. Through the implementation of three distinct neural SDEs models—Langevin-type SDE, Linear Noise SDE, and Geometric SDE—the researchers demonstrated stable and exceptional performance across interpolation, prediction, and classification tasks, even in the presence of data drift.

Traditionally, addressing data drift necessitated labor-intensive and costly engineering adjustments to adapt to evolving data landscapes. However, the team’s methodology offers a proactive solution by ensuring AI models remain resilient to data drift from the outset, obviating the need for extensive relearning processes.

Professor Lim underscored the study’s significance in fortifying the resilience of time series AI models against dynamic data environments, thereby enabling practical applications across diverse industries. Lead author YongKyung Oh highlighted the team’s dedication to advancing technologies for monitoring time series data drift and reconstructing data, paving the way for widespread adoption by Korean enterprises.

This groundbreaking research has been recognized as a top 5% spotlight paper at the prestigious International Conference on Learning Representations (ICLR), underscoring its global impact. Supported by various institutions, such as the Korea Health Industry Development Institute (KHIDI) and the National Research Foundation of Korea (NRF), this study represents a significant advancement in the field of AI and data science.

Journal Reference
YongKyung Oh, Dongyoung Lim, Sungil Kim, Stable Neural Stochastic Differential Equations in Analyzing Irregular Time Series Data, selected as ICLR 2024 Spotlight, (2024).