Research map / Time series / Long-term forecasting
From Similarity to Superiority: Channel Clustering for 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 a plug-and-play Channel Clustering Module (CCM) that groups time series channels by intrinsic similarity via learned prototypes and cluster-aware feed-forward layers, improving CI and CD forecasters on long-term, short-term (M4, new stock dataset) and zero-shot forecasting.
Problem
Channel-independent (CI) forecasters generalize poorly to unseen instances and ignore cross-channel interactions, while channel-dependent (CD) models oversmooth and fit individual channels poorly; no balanced channel strategy exists for time series forecasting.
Approach
Cluster channels by pairwise RBF-kernel similarity using learned cluster embeddings and a sparse cross-attention that yields prototypes; then use one feed-forward per cluster (weighted by soft membership) so cluster identity replaces channel identity, with a self-supervised ClusterLoss that favors intra-cluster similarity and inter-cluster separation.
Key results
- CCM improves long-term MSE in 90.27% of 144 settings; average IMP of about 2.4% across base models per the abstract.
- On M4 averaged over sub-datasets, CCM with DLinear beats DLinear by 11.62% and TimesNet by 8.88% (as stated in the text).
- Zero-shot ETT transfer: CCM improves base models in 48 scenarios, with DLinear and PatchTST average improvements of 10.48% and 11.13%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CCM to TSMixer | TSMixer · Long-term forecasting across 9 datasets and 144 settings | MSE: 90.27% | Helped |
| Add CCM to CI models | DLinear and PatchTST · Zero-shot forecasting on ETT datasets | MSE: 10.48%; 11.13% | Helped |
| Add CCM to univariate forecasting models | M4 · M4 sub-datasets | OWA: 11.62%; 8.88% | Helped |
| Add CCM to forecasting models | Long-term forecasting models · Long-term forecasting | MSE: 2.4% | Helped |
| Add CCM to forecasting models | Short-term forecasting models · Short-term forecasting | OWA: 7.2% | Helped |
| Add CCM instead of PRReg | CI/CD Linear and Transformer models · Long-term forecasting | MSE: 13 out of 16 settings | Mixed |
| Vary the cluster ratio (number of clusters divided by number of channels) (ablation) | CCM with TSMixer, DLinear, PatchTST, and TimesNet · ETTh1 and ETTm1 (horizon 96) | MSE | Mixed |
| Use channel clustering with any number of clusters (ablation) | DLinear and PatchTST with CCM · ETTh1 and ETTm1 (horizon 96) | MSE | Helped |
Surprising findings
- Randomly shuffling channel order (removing channel identity) hurts all four base models, including CI models, and the channel-identity reliance is anti-correlated with channel similarity (PCC around -0.5 to -0.7).
Limitations
- Clustering quality depends on the choice of similarity measure; Euclidean distance with fixed sigma=5 is used as a proxy.
- Stock dataset is newly constructed and not released, limiting external validation.
- Zero-shot forecasting on multivariate data with CD models requires the same number of channels as training.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Exchange ILI Weather Electricity Traffic M4 Stock (new)
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Exchange ILI Weather Electricity Traffic M4 Stock (new)
Metrics: MSE MAE SMAPE MASE OWA
Baselines: TSMixer DLinear PatchTST TimesNet PRReg
Builds on: PatchTST DLinear TSMixer TimesNet RevIN
Closest papers by meaning
- CSformer: Combining Channel Independence and Mixing for Robust Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.90)
- TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting (ICML 2025 · similarity 0.89)
- Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators (ICLR 2024 · similarity 0.89)
- SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion (NeurIPS 2024 · similarity 0.88)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.88)
- HyperMixer: Specializable Hypergraph Channel Mixing for Long-term 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.