Research map / Time series / Long-term forecasting
Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach
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 T-PatchGNN, which splits each univariate irregular series into time-horizon-aligned transformable patches and models inter-series correlation with time-varying adaptive graph neural networks for IMTS forecasting.
Problem
Forecasting irregular multivariate time series (IMTS), where sampling is irregular and observations are asynchronous across variables; existing methods rely on ODEs and pre-alignment that causes sequence length explosion and poorly models inter-series correlations.
Approach
Patch each univariate irregular series independently by fixed time windows (variable observation count per patch) so patches are time-aligned across variables without pre-alignment; encode patches with a transformable time-aware convolution, model intra-series dependencies with a Transformer, and inter-series correlations with GNNs over learned time-varying adaptive graphs.
Key results
- T-PatchGNN achieves best MSE/MAE on all four datasets, e.g., PhysioNet MSE 4.98e-3 vs 5.59e-3 for the next best (GRU-D).
- Outperforms the second-best baseline by over 10% per the paper's claim.
- Achieves at least 65x faster training and 15x faster inference than ODE-based IMTS forecasting models on MIMIC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add transformable patching instead of canonical pre-alignment (ablation) | T-PATCHGNN · PhysioNet, MIMIC, Human Activity, USHCN | MSE: 5.27 ± 0.06 -> 4.98 ± 0.08; 1.73 ± 0.02 -> 1.69 ± 0.03; 2.85 ± 0.05 -> 2.66 ± 0.03; 5.32 ± 0.02 -> 5.00 ± 0.04 | Helped |
| Use transformable patching rather than standard time-series patching (ablation) | T-PATCHGNN · PhysioNet, MIMIC, Human Activity, USHCN | MSE: 5.64 ± 0.13 -> 4.98 ± 0.08; 1.72 ± 0.01 -> 1.69 ± 0.03; 2.82 ± 0.02 -> 2.66 ± 0.03; 5.20 ± 0.07 -> 5.00 ± 0.04 | Mixed |
| Add variable embeddings to adaptive graph construction (ablation) | T-PATCHGNN · PhysioNet, MIMIC, Human Activity, USHCN | MSE: 10.53 ± 0.15 -> 4.98 ± 0.08; 3.71 ± 0.02 -> 1.69 ± 0.03; 2.79 ± 0.03 -> 2.66 ± 0.03; 5.08 ± 0.03 -> 5.00 ± 0.04 | Helped |
| Add patch embeddings to adaptive graph construction (ablation) | T-PATCHGNN · PhysioNet, MIMIC, Human Activity, USHCN | MSE: 5.10 ± 0.02 -> 4.98 ± 0.08; 1.71 ± 0.01 -> 1.69 ± 0.03; 2.79 ± 0.02 -> 2.66 ± 0.03; 5.08 ± 0.02 -> 5.00 ± 0.04 | Helped |
| Add the Transformer module (ablation) | T-PATCHGNN · PhysioNet, MIMIC, Human Activity, USHCN | MSE: 5.10 ± 0.12 -> 4.98 ± 0.08; 1.71 ± 0.02 -> 1.69 ± 0.03; 2.78 ± 0.04 -> 2.66 ± 0.03; 5.22 ± 0.06 -> 5.00 ± 0.04 | Helped |
Surprising findings
- Standard (regular) time series patching can perform worse than no patching on PhysioNet, since patch temporal resolution varies with sampling density.
Limitations
- Evaluated only on four datasets from healthcare, biomechanics, and climate science.
- Patch window size is a sensitive hyperparameter that must be tuned per dataset.
Experimental setup
Datasets: PhysioNet MIMIC Human Activity USHCN
Benchmarks: PhysioNet MIMIC Human Activity USHCN
Metrics: MSE MAE
Baselines: DLinear TimesNet PatchTST Crossformer Graph WaveNet MTGNN StemGNN CrossGNN FourierGNN GRU-D SeFT RainDrop
Builds on: TTCN PatchTST Graph WaveNet MTGNN Neural ODE
Closest papers by meaning
- Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series (ICML 2025 · similarity 0.91)
- HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting (ICML 2025 · similarity 0.89)
- Time-IMM: A Dataset and Benchmark for Irregular Multimodal Multivariate Time Series (NeurIPS 2025 · similarity 0.87)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.87)
- Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs (ICLR 2025 · similarity 0.87)
- MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · 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.