Research map / Time series / Long-term forecasting
TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis
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 TimeCHEAT, a Transformer for irregularly sampled multivariate time series that applies channel-dependent embedding locally within patches (via bipartite graph edge prediction) and channel-independent attention globally across patches.
Problem
Channel strategy for irregularly sampled multivariate time series (ISMTS): channel-independent modeling suffers from sampling sparsity and channel-dependent modeling loses diversity, and prior work treats CI and CD as mutually exclusive global choices.
Approach
Combine CD locally (patch-level bipartite graph with GNN that learns reference-point edge weights to produce fixed-length embeddings without timing-interval priors) and CI globally (shared Transformer applied per channel across patches), so the two channel strategies are used at different scales.
Key results
- On P19 classification, TimeCHEAT reaches 89.5 AUROC and 56.1 AUPRC, outperforming the baselines (best prior ViTST 89.2 AUROC).
- On PAM 8-class classification, TimeCHEAT achieves 96.5 accuracy and 97.0 F1, versus ViTST's 95.8 accuracy.
- On P12 classification TimeCHEAT gets 84.5 AUROC, trailing ViTST's 85.1 by about 0.5 points; in forecasting it is competitive, with GraFITi best on most datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use local channel correlation in embedding learning instead of removing it (ablation) | TimeCHEAT · P19 | AUROC: 88.0 ± 3.1 -> 89.5 ± 1.9 | Helped |
| Use local channel correlation in embedding learning instead of removing it (ablation) | TimeCHEAT · P12 | AUROC: 83.5 ± 0.9 -> 84.5 ± 0.7 | Helped |
| Use the channel-independent Transformer encoder instead of iTransformer (ablation) | TimeCHEAT · P19 | AUROC: 87.6 ± 2.4 -> 89.5 ± 1.9 | Helped |
| Use the channel-independent Transformer encoder instead of iTransformer (ablation) | TimeCHEAT · P12 | AUROC: 79.9 ± 1.5 -> 84.5 ± 0.7 | Helped |
| Use both local channel correlation and the channel-independent Transformer encoder rather than removing correlation and using iTransformer (ablation) | TimeCHEAT · P19 | AUROC: 86.8 ± 2.7 -> 89.5 ± 1.9 | Helped |
| Use both local channel correlation and the channel-independent Transformer encoder rather than removing correlation and using iTransformer (ablation) | TimeCHEAT · P12 | AUROC: 80.1 ± 1.7 -> 84.5 ± 0.7 | Helped |
| Use local graph embedding instead of CI mTAND embedding (ablation) | TimeCHEAT · P19 | AUROC: 87.4 ± 2.3 -> 89.5 ± 1.9 | Helped |
| Use local graph embedding instead of CI mTAND embedding (ablation) | TimeCHEAT · P12 | AUROC: 84.3 ± 0.8 -> 84.5 ± 0.7 | No clear effect |
Surprising findings
- The paper reports that a fully CI model (CI embedding plus CI Transformer) yields nearly the worst accuracy among tested conditions, contrary to the expectation that CI is sufficient.
Limitations
- Forecasting results are not uniformly best; GraFITi outperforms TimeCHEAT on several datasets.
- On P12 classification it trails ViTST slightly.
Experimental setup
Datasets: P19 P12 PAM PhysioNet USHCN MIMIC-III MIMIC-IV Physionet12
Benchmarks: P19 P12 PAM PhysioNet USHCN MIMIC-III MIMIC-IV Physionet12
Metrics: AUROC AUPRC accuracy precision recall F1 score MSE
Baselines: Transformer Trans-mean GRU-D SeFT mTAND IP-Net DGM2-O MTGNN Raindrop ViTST RNN-VAE L-ODE-RNN
Builds on: mTAND PatchTST GraFITi Graph Attention Network
Closest papers by meaning
- IMTS is Worth Time $\times$ Channel Patches: Visual Masked Autoencoders for Irregular Multivariate Time Series Prediction (ICML 2025 · similarity 0.90)
- CSformer: Combining Channel Independence and Mixing for Robust Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.89)
- MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · similarity 0.88)
- SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion (NeurIPS 2024 · similarity 0.88)
- From Similarity to Superiority: Channel Clustering for Time Series Forecasting (NeurIPS 2024 · similarity 0.88)
- TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting (ICML 2025 · similarity 0.88)
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.