Search papers, labs, and topics across Lattice.
This paper introduces Harmonic Dataset Distillation for Time Series Forecasting (HDT), a novel dataset distillation method tailored for time series data. HDT leverages Fast Fourier Transform (FFT) to decompose time series into sinusoidal bases and aligns periodic structures via Harmonic Matching in the frequency domain. Experiments demonstrate that HDT achieves strong cross-architecture generalization and scalability, making it practical for large-scale applications.
Dataset distillation for time series can now be done efficiently and scalably by operating in the frequency domain, sidestepping the limitations of prior methods.
Time Series forecasting (TSF) in the modern era faces significant computational and storage cost challenges due to the massive scale of real-world data. Dataset Distillation (DD), a paradigm that synthesizes a small, compact dataset to achieve training performance comparable to that of the original dataset, has emerged as a promising solution. However, conventional DD methods are not tailored for time series and suffer from architectural overfitting and limited scalability. To address these issues, we propose Harmonic Dataset Distillation for Time Series Forecasting (HDT). HDT decomposes the time series into its sinusoidal basis through the FFT and aligns the core periodic structure by Harmonic Matching. Since this process operates in the frequency domain, all updates during distillation are applied globally without disrupting temporal dependencies of time series. Extensive experiments demonstrate that HDT achieves strong cross-architecture generalization and scalability, validating its practicality for large-scale, real-world applications.
