Research map / Time series / Long-term forecasting
HyperIMTS: Hypergraph Neural Network for Irregular 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 HyperIMTS, a hypergraph neural network that represents each observed value as a node linked by temporal and variable hyperedges, enabling padding-free irregular multivariate time series forecasting with irregularity-aware variable message passing.
Problem
Forecasting irregular multivariate time series (IMTS), where observation times differ within and across variables; existing padding-based methods inflate data and bipartite/set-based methods fail to capture dependencies among unaligned observations.
Approach
Represent original observations (no padding) as hypergraph nodes connected by temporal and variable hyperedges, and learn variable dependencies via hyperedge-to-hyperedge attention whose similarity mixes time-aligned observation similarity and overall variable-embedding similarity, weighted by the fraction of shared timestamps.
Key results
- HyperIMTS achieves best MSE on four of five datasets, e.g. 0.4259 on MIMIC-III and 0.2174 on MIMIC-IV versus GraFITi's 0.4534 and 0.2454.
- HyperIMTS gives up to 11.4% MSE improvement over the next-best model GraFITi.
- Ablations show removing inter-variable irregularity-aware dependencies (w/o IAVD) or variable hyperedges (w/o VE) increases MSE, with w/o VE raising Human Activity MSE from 0.0421 to 1.3855.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add variable hyperedges (ablation) | HyperIMTS · MIMIC-III, MIMIC-IV, PhysioNet'12, Human Activity, USHCN | MSE: 0.9556 $\pm$ 0.0029 -> 0.4259 $\pm$ 0.0021; 0.6293 $\pm$ 0.0053 -> 0.2174 $\pm$ 0.0009; 0.6945 $\pm$ 0.0002 -> 0.2996 $\ | Helped |
| enable irregularity-aware variable dependencies and hyperedge-to-hyperedge message passing (ablation) | HyperIMTS · MIMIC-III, MIMIC-IV, PhysioNet'12, Human Activity, USHCN | MSE: 0.4466 $\pm$ 0.0007 -> 0.4259 $\pm$ 0.0021; 0.2358 $\pm$ 0.0012 -> 0.2174 $\pm$ 0.0009; 0.3050 $\pm$ 0.0006 -> 0.2996 $\ | Helped |
| use irregularity-aware variable dependencies instead of overall variable dependencies only (ablation) | HyperIMTS · MIMIC-III, MIMIC-IV, PhysioNet'12, Human Activity, USHCN | MSE: 0.4317 $\pm$ 0.0030 -> 0.4259 $\pm$ 0.0021; 0.2189 $\pm$ 0.0024 -> 0.2174 $\pm$ 0.0009; 0.3014 $\pm$ 0.0003 -> 0.2996 $\ | Helped |
| add temporal hyperedges (ablation) | HyperIMTS · MIMIC-III, MIMIC-IV, PhysioNet'12, Human Activity, USHCN | MSE: 0.4954 $\pm$ 0.0043 -> 0.4259 $\pm$ 0.0021; 0.2652 $\pm$ 0.0005 -> 0.2174 $\pm$ 0.0009; 0.3029 $\pm$ 0.0003 -> 0.2996 $\ | Helped |
| use learnable temporal hyperedges instead of non-learnable sinusoidal embeddings (ablation) | HyperIMTS · MIMIC-III, MIMIC-IV, PhysioNet'12, Human Activity, USHCN | MSE: 0.4403 $\pm$ 0.0024 -> 0.4259 $\pm$ 0.0021; 0.2333 $\pm$ 0.0004 -> 0.2174 $\pm$ 0.0009; 0.3029 $\pm$ 0.0003 -> 0.2996 $\ | Helped |
Surprising findings
- On USHCN, HyperIMTS (0.1738 MSE) is not best; Warpformer (0.1565) is, and the paper notes high variance on that dataset.
Limitations
- Does not support multimodal inputs such as clinical text notes or images.
- Attention-based computation is more resource-intensive than state-space approaches.
- Results on USHCN show high variance and HyperIMTS is not best there.
Experimental setup
Datasets: MIMIC-III MIMIC-IV PhysioNet'12 Human Activity USHCN
Benchmarks: MIMIC-III MIMIC-IV PhysioNet'12 Human Activity USHCN
Metrics: MSE MAE training time per iteration GPU memory
Baselines: GraFITi tPatchGNN Warpformer GRU-D Raindrop mTAN PatchTST Crossformer iTransformer Informer MOIRAI PrimeNet
Builds on: hypergraph neural networks GraFITi tPatchGNN Transformer multi-head attention
Closest papers by meaning
- HyperMixer: Specializable Hypergraph Channel Mixing for Long-term Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.90)
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.89)
- Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series (ICML 2025 · similarity 0.89)
- TimePro: Efficient Multivariate Long-term Time Series Forecasting with Variable- and Time-Aware Hyper-state (ICML 2025 · similarity 0.86)
- Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
- HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · 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.