Wavelet flow matching improves multichannel time series synthesis
Wavelet Flow Matching for Time Series
Machine LearningArtificial Intelligence
Summary
Synthetic time series data is used for various important tasks but is hard to make realistic, especially when showing patterns that happen over many time scales and across multiple signals. The authors design a method that first breaks down the series into different levels of detail using a wavelet transform, then models each level separately, capturing both broad patterns and fine details naturally. They also use a special transformer model to understand and create realistic links between multiple related time signals. Their method performs very well compared to others on multiple benchmark tests.
What this means in practice
- •For privacy engineers: Generate realistic synthetic multivariate time series for privacy-safe data sharing that preserves essential temporal and cross-signal patterns.
- •For machine learning engineers: Create diverse synthetic time series data to augment training datasets improving robustness of downstream models handling multichannel sequential data.
Authors
Lucas Poinsignon, Jorge da Silva Gonçalves, Samuel Ruipérez-Campillo, Julia E. Vogt
Abstract
Synthetic time series are increasingly used for data augmentation, privacy-preserving data sharing, and downstream model development, yet faithfully reproducing both multi-scale temporal structure and cross-channel dependencies remains challenging. We study multivariate time-series generation through flow matching in the wavelet domain. By operating on multilevel discrete wavelet coefficients rather than directly in the time domain, the model represents coarse structure and progressively finer details at separate scales. Their naturally different variances further induce an implicit coarse-to-fine generative process without requiring an explicit multi-scale schedule. Since the transform acts independently on each channel, we pair it with a channel-token transformer whose attention directly models cross-channel dependencies. Across seven benchmark datasets and four sequence lengths, our method is best or tied on a majority of dataset-metric combinations, with the largest and most consistent improvements in Context-FID and discriminative score.