Research map / Time series / Long-term forecasting
HyperMixer: Specializable Hypergraph Channel Mixing for Long-term Multivariate Time Series Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HyperMixer, a plug-in that models higher-order, time-varying channel correlations in long-term multivariate forecasting via a channel hypergraph with mixture-of-experts-style specializable hypergraph structures.
Problem
Long-term multivariate time series forecasting needs channel correlations, but existing methods treat channels as independent or model only pairwise interactions with a shared network, missing higher-order group interactions and time-varying correlations.
Approach
Organize channels into a hypergraph and run two-phase message passing (channel-to-group, group-to-channel); a router with multiple hypergraph-structure experts (plus a shared expert) produces period-specific hypergraph structures, and the module is inserted as a plugin into existing backbones.
Key results
- HyperMixer improves the backbones by an average of 3.46% across four backbones and seven datasets (Table 2).
- With PatchTST on Traffic, MSE drops from 0.396 to 0.381 with HyperMixer, versus 0.389 with CCM.
- Ablation on TimesNet: HCM gives 1.87% gain on ETTh1 and HS2L adds 1.64% on Weather, indicating both components contribute.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HyperMixer to PatchTST | PatchTST · ETTh1 (four prediction lengths averaged) | MSE: 0.412 -> 0.408 | Helped |
| Add HyperMixer to PatchTST | PatchTST · ETTh2 (four prediction lengths averaged) | MSE: 0.330 -> 0.327 | Helped |
| Add HyperMixer to PatchTST | PatchTST · ETTm1 (four prediction lengths averaged) | MSE: 0.353 -> 0.346 | Helped |
| Add HyperMixer to PatchTST | PatchTST · ETTm2 (four prediction lengths averaged) | MSE: 0.262 -> 0.256 | Helped |
| Add HyperMixer to PatchTST | PatchTST · Weather (four prediction lengths averaged) | MSE: 0.225 -> 0.223 | Helped |
| Add HyperMixer to PatchTST | PatchTST · Traffic (four prediction lengths averaged) | MSE: 0.389 -> 0.381 | Helped |
| Add HCM, without HS2L (ablation) | TimesNet · ETTh1 (four prediction lengths averaged) | MSE: 0.458 -> 0.451 | Helped |
| Add HS2L to TimesNet+HCM (ablation) | TimesNet · ETTh1 (four prediction lengths averaged) | MSE: 0.451 -> 0.447 | Helped |
| Add HCM, without HS2L (ablation) | TimesNet · ETTm2 (four prediction lengths averaged) | MSE: 0.291 -> 0.290 | Helped |
| Add HS2L to TimesNet+HCM (ablation) | TimesNet · ETTm2 (four prediction lengths averaged) | MSE: 0.290 -> 0.286 | Helped |
| Add HCM, without HS2L (ablation) | TimesNet · Weather (four prediction lengths averaged) | MSE: 0.259 -> 0.254 | Helped |
| Add HS2L to TimesNet+HCM (ablation) | TimesNet · Weather (four prediction lengths averaged) | MSE: 0.254 -> 0.249 | Helped |
Surprising findings
- Ablation shows HS2L gains are larger on datasets with minute-level sampling (Weather, ETTm2), attributed to richer time-varying patterns.
Limitations
- Hyperparameters (number of experts, hyperedges) must be tuned per dataset and backbone.
- Results averaged over four prediction lengths; TimeMixer CCM numbers are from the authors' own reproduction.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic
Metrics: MSE MAE
Baselines: CCM PatchTST TimeMixer DLinear TimesNet
Builds on: hypergraph neural networks mixture of experts PatchTST TimeMixer DLinear
Closest papers by meaning
- HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting (ICML 2025 · similarity 0.90)
- Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting (NeurIPS 2024 · similarity 0.88)
- xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories (NeurIPS 2025 · similarity 0.88)
- From Similarity to Superiority: Channel Clustering for Time Series Forecasting (NeurIPS 2024 · similarity 0.87)
- TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · similarity 0.87)
- CSformer: Combining Channel Independence and Mixing for Robust Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.87)
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.