Latent Flow Matching Method Aligns Spectra for Time Series Generation
Time series generation with spectrally aligned latent flow matching
The method outperforms a base latent-flow baseline and prior state of the art on realness and efficiency metrics, making synthetic time series more dependable as training data for forecasting or anomaly-detection work.
A paper on arXiv (2609.21989) proposes a spectrally aligned latent-flow generator for time series, training the flow-matching latent space to preserve dynamical properties of the source data.
The method adds fine-tuning losses built on canonical signal representations, including the Fourier, wavelet and signature transforms, to fix spectral mismatch artefacts caused by latent compression.
These interpretable transformations let the authors verify that synthetic signals match real ones on features such as smoothness and targeted spectral content, rather than relying on pointwise reconstruction losses alone.
Against a base latent-flow model and the state of the art on real-world long-range univariate and multivariate benchmark datasets, the aligned models score better on metrics of signal realness and computational efficiency.
Time series generation with spectrally aligned latent flow matching
Latent flow models have proven to be a reliable and cost-effective method for time series generation. However, the latent compression induces unwanted artefacts, such as a spectral mismatch with respect to the underlying dataset, thus hindering their use as training surrogates. In this article, we propose a spectrally-aligned latent-flow time series generator, where the latent space for flow matching is trained to preserve dynamical properties that are relevant for the suitability of synthetic samples. We find that incorporating fine-tuning losses based on canonical signal representations such as the Fourier, wavelet and signature transforms helps overcome these issues. The interpretability of these transformations allows us to ensure that the synthetic signals are aligned with the true ones in terms of relevant features, such as smoothness or targeted spectral content, as opposed to relying on pointwise reconstruction losses only. We compare the proposed aligned models against a base latent-flow model and the state of the art over real-world long-range univariate and multivariate benchmark datasets. Our quantitative results validate the superiority of the proposed method in terms of its performance on metrics reflecting signal realness and computational efficiency, while being aligned to the training set with respect to its local structure.
This report is categorized under AITOP's "Papers" desk. Key entities and technical topics under continuous tracking include Flow Matching, time-series generation, synthetic data.
Source attribution: curated from arXiv cs.LG (@Camilo Carvajal Reyes, Felipe Tobar).