Harmonic Dataset Distillation for Time Series Forecasting

2026-03-04

AI Summary

This work proposes a novel time series dataset distillation approach that overcomes the limitations of existing methods—namely, architectural overfitting, poor scalability, and inadequate preservation of temporal dependencies. For the first time, harmonic matching in the frequency domain is introduced to time series distillation. By decomposing original sequences into sinusoidal bases via the Fast Fourier Transform, the method aligns essential periodic structures in the frequency domain to achieve global optimization, thereby constructing compact yet highly effective synthetic datasets. This strategy inherently preserves temporal dependencies and significantly enhances cross-model generalization and scalability. Extensive experiments on multiple real-world time series datasets demonstrate that the proposed method consistently outperforms current distillation techniques across diverse forecasting architectures.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Planning, Routing, and Scheduling: Optimization of Spatio-temporal Systems
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

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.

Problem

• Time Series Forecasting
• Dataset Distillation
• Computational Cost
• Scalability
• Architectural Overfitting

Innovation

• Dataset Distillation
• Time Series Forecasting
• Frequency Domain
• Harmonic Matching
• FFT