Research map / Time series / Long-term forecasting
Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series
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 Hi-Patch, a hierarchical patch graph network that models irregular multivariate time series at multiple scales via intra-patch and inter-patch GNN layers, evaluated on IMTS forecasting and classification across 8 datasets.
Problem
Multi-scale modeling of irregular multivariate time series (IMTS), where variables have distinct original sampling rates and asynchronous observations, which existing downsampling/segmentation-based multi-scale methods cannot handle.
Approach
Encode each observation as a graph node, build fully connected intra-patch graphs for densely sampled variables, aggregate nodes into patch-level nodes, and stack inter-patch graph layers that progressively merge patches (P, 2P, 4P, ...) so that scale-specific GAT updates capture fine-to-coarse temporal and inter-variable dependencies without interpolation.
Key results
- Hi-Patch achieves the best performance in 62 out of 72 metrics across 8 datasets for classification and forecasting tasks.
- On Human Activity forecasting, Hi-Patch reaches MSE 2.57e-3 versus 2.61e-3 for the next best (Warpformer).
- On P19 classification, Hi-Patch achieves AUROC 92.1 vs 91.8 for Warpformer; on MIMIC-III AUPRC 57.3 vs 52.5-ish baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hierarchical multi-scale architecture rather than extracting only original-scale features (ablation) | Hi-Patch · Human Activity | MSE×10−3: 2.70 ± 0.04 -> 2.57 ± 0.02 | Helped |
| use hierarchical multi-scale architecture rather than extracting only original-scale features (ablation) | Hi-Patch · USHCN | MSE×10−1: 5.35 ± 0.13 -> 4.94 ± 0.05 | Helped |
| retain asynchronous edges between different variables at different times (ablation) | Hi-Patch · Human Activity | MSE×10−3: 2.68 ± 0.04 -> 2.57 ± 0.02 | Helped |
| retain asynchronous edges between different variables at different times (ablation) | Hi-Patch · USHCN | MSE×10−1: 5.21 ± 0.03 -> 4.94 ± 0.05 | Helped |
| use separate attention parameter sets for the three edge types (ablation) | Hi-Patch · Human Activity | MSE×10−3: 2.60 ± 0.01 -> 2.57 ± 0.02 | No clear effect |
| use separate attention parameter sets for the three edge types (ablation) | Hi-Patch · USHCN | MSE×10−1: 4.99 ± 0.03 -> 4.94 ± 0.05 | Helped |
| aggregate nodes with multi-time attention instead of mean aggregation (ablation) | Hi-Patch · Human Activity | MSE×10−3: 2.77 ± 0.02 -> 2.57 ± 0.02 | Helped |
| aggregate nodes with multi-time attention instead of mean aggregation (ablation) | Hi-Patch · USHCN | MSE×10−1: 5.23 ± 0.04 -> 4.94 ± 0.05 | Helped |
Surprising findings
- On PhysioNet and MIMIC-III, using too many scales performs worse than a single scale because these samples are short-length, long-span and lack fine-grained features.
Limitations
- Intra-patch graph layer has quadratic complexity in observations per patch, limiting scalability to dense data.
- Scale choice (patch size) is manually tuned and the number of scales can hurt performance on short-length, long-span samples.
Experimental setup
Datasets: PhysioNet MIMIC-III Human Activity USHCN P19 P12
Benchmarks: PhysioNet MIMIC-III Human Activity USHCN P19 P12
Metrics: MSE MAE AUROC AUPRC
Baselines: iTransformer ModernTCN TimesNet PatchTST Pathformer TimeMixer MSGNet Warpformer tPatchGNN GraFITi
Builds on: tPatchGNN mTAND Warpformer graph attention network PatchTST
Closest papers by meaning
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.91)
- HyperIMTS: Hypergraph Neural Network for Irregular Multivariate Time Series Forecasting (ICML 2025 · similarity 0.89)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.87)
- Hierarchical Classification Auxiliary Network for Time Series Forecasting (AAAI 2025 · similarity 0.86)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.86)
- MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · similarity 0.86)
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.