Research map / Time series / Long-term forecasting
Harmonic Dataset Distillation for Time Series Forecasting
Seungha Hong, Sanghwan Jang, Wonbin Kweon, Suyeon Kim, Gyuseok Lee, Hwanjo Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sinusoidal decomposition (frequency-domain updates) to gradient matching (ablation) | HDT, DLinear · ETTh1 (M=384, avg over evaluation models) | MSE: 0.583 -> 0.545 | Improved |
| add Harmonic Matching on top of decomposition (ablation) | HDT, DLinear · ETTh1 (M=384, avg over evaluation models) | MSE: 0.545 -> 0.420 | Improved |
| add Harmonic Matching on top of decomposition (ablation) | HDT, DLinear · Traffic (M=384, avg over evaluation models) | MSE: 0.902 -> 0.760 | Improved |
| add Harmonic Matching on top of decomposition (ablation) | HDT, DLinear · Electricity (M=384, avg over evaluation models) | MSE: 0.376 -> 0.226 | Improved |
| add sinusoidal decomposition to gradient matching (ablation) | HDT, DLinear · ETTm2 (M=384, avg over evaluation models) | MSE: 0.402 -> 0.325 | Improved |
| use HDT instead of CondTSF with xPatch backbone evaluated on DLinear (ablation) | distillation method · ETTh1 (M=384) | MSE: 1.172 -> 0.412 | Improved |
| use HDT instead of CondTSF | distillation method, DLinear backbone · ETTh1 (M=384, evaluated on DLinear) | MSE: 0.510 -> 0.430 | Improved |
| use HDT instead of CondTSF | distillation method, DLinear backbone · Traffic (M=384, evaluated on DLinear) | MSE: 0.716 -> 0.679 | Improved |
| use HDT instead of CondTSF on large CA dataset | distillation method · CA (LargeST, M=384) | MSE: 197.95 -> 46.63 | Improved |
| train iTransformer on distilled data instead of full data | iTransformer · Electricity | training time (s): 1650.33 -> 1.98 | Improved |
| fine-tune Moirai-Large on HDT few-shot data instead of full fine-tuning | Moirai-Large · ETTh1 | MSE: 1.383 -> 1.417 | Mixed |
| use HDT instead of CondTSF | distillation runtime, DLinear · ETTh1 (100 outer loops) | runtime (s): 32.13 -> 36.37 | Hurt |
Closest papers by meaning
- HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.85)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.84)
- CondTSF: One-line Plugin of Dataset Condensation for Time Series Forecasting (NeurIPS 2024 · similarity 0.84)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.84)
- HDMixer: Hierarchical Dependency with Extendable Patch for Multivariate Time Series Forecasting (AAAI 2024 · similarity 0.83)
- DDN: Dual-domain Dynamic Normalization for Non-stationary Time Series Forecasting (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
