Research map / Time series / Long-term forecasting
Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer 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 Ada-MSHyper, an adaptive multi-scale hypergraph transformer that learns data-driven hypergraphs per scale, adds node and hyperedge constraint losses, and models intra- and inter-scale group-wise interactions for long-range time series forecasting.
Problem
Transformers for time series forecasting struggle with sparse semantic information per time point (pair-wise attention is insufficient) and with entangled temporal variations within and across scales.
Approach
Learns adaptive (data-driven) multi-scale hypergraph incidence matrices to model group-wise interactions, adds node constraint (semantic similarity clustering of nodes to hyperedges) and hyperedge constraint (cosine/Euclidean-based loss on hyperedge representations) to separate temporal variations, and uses intra-scale hypergraph attention plus inter-scale hyperedge attention.
Key results
- Long-range multivariate forecasting: Ada-MSHyper reports best MSE/MAE on all six averaged benchmarks (e.g., Traffic 0.415/0.262, ETTh1 0.418/0.426), with a claimed 4.67% MSE and 3.47% MAE average reduction vs best baseline.
- Short-range PEMS: claimed average error reduction of 13.72% MSE and 3.59% MAE vs best baseline (e.g., PEMS04 0.080/0.189).
- Ablation on ETTh1 shows removing the NHC mechanism (w/o NHC) worsens results versus full model (e.g., horizon 96 MSE 0.393 vs 0.372).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive hypergraph learning instead of adaptive graph learning (ablation) | Ada-MSHyper · ETTh1, prediction length 96 | MSE: 0.542 -> 0.372 | Helped |
| Use multi-scale incidence matrices instead of one incidence matrix (ablation) | Ada-MSHyper · ETTh1, prediction length 96 | MSE: 0.422 -> 0.372 | Helped |
| Use adaptive hypergraphs instead of predefined multi-scale hypergraphs (ablation) | Ada-MSHyper · ETTh1, prediction length 96 | MSE: 0.386 -> 0.372 | Helped |
| Add node constraint (ablation) | Ada-MSHyper · ETTh1, prediction length 96 | MSE: 0.390 -> 0.372 | Helped |
| Add hyperedge constraint (ablation) | Ada-MSHyper · ETTh1, prediction length 96 | MSE: 0.384 -> 0.372 | Helped |
| Add node and hyperedge constraints (ablation) | Ada-MSHyper · ETTh1, prediction length 96 | MSE: 0.393 -> 0.372 | Helped |
Limitations
- Paper has many placeholders (xxx) in abstract and appendix, and inconsistent dataset counts (5, 7, or 11 benchmarks).
- Baselines were re-run with varying input/output settings; fairness across baselines is unclear.
- Node and hyperedge constraint loss formulations contain notational errors (e.g., L_node normalization), making exact replication harder.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic PEMS03 PEMS04 PEMS07 PEMS08
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic PEMS03 PEMS04 PEMS07 PEMS08
Metrics: MSE MAE
Baselines: iTransformer MSHyper TimeMixer PatchTST Crossformer TimesNet DLinear FiLM FEDformer Autoformer SCINet WITRAN
Builds on: MSHyper HGNN PatchTST iTransformer
Closest papers by meaning
- HyperMixer: Specializable Hypergraph Channel Mixing for Long-term Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.88)
- Deep Representation Learning for Forecasting Recursive and Multi-Relational Events in Temporal Networks (AAAI 2025 · similarity 0.87)
- Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting (AAAI 2025 · similarity 0.86)
- HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting (ICML 2025 · similarity 0.86)
- Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting (ICLR 2024 · similarity 0.86)
- Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting (ICML 2024 · similarity 0.85)
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.