A1187
Title: Structure-preserving time series synthesis
Authors: Changwook Lee - Yonsei University (Korea, South)
Joseph Hyun Tae Kim - Yonsei University (Korea, South)
Nathaniel Kang - Kyungpook National University (Korea, South)
Jongho Im - Yonsei University (Korea, South) [presenting]
Abstract: Synthetic time series are increasingly important in industrial and operational statistics, where long-horizon data are needed for monitoring, forecasting, planning, and privacy-conscious data sharing. However, existing deep generative models often struggle to preserve realistic temporal dynamics and to scale effectively to long sequences. Rather than increasing architectural complexity, structural properties that characterize realistic time series in industrial settings are prioritized, distinguishing temporal patterns from cross-sectional structure. From this perspective, LAMUS (Latent MixUp-based Synthesis), a computationally efficient framework for synthetic time series generation, is introduced. LAMUS generates new temporal patterns through mixup in a low-dimensional Latent space and restores cross-sectional structure through calibration. The Cluster Score (CLS), a new evaluation metric designed to assess temporal pattern preservation at both the sample and distributional levels, is also proposed. Experiments on diverse real-world benchmarks, including energy, finance, transportation, and healthcare data, show that LAMUS consistently outperforms state-of-the-art generative models across key quality measures while maintaining strong scalability to long sequences. These results suggest that structural alignment, rather than architectural complexity, is central to high-fidelity and practically useful synthetic time series generation for industrial statistics.
