Research map / Time series / Long-term forecasting
Graph-based multivariate forecasting
Forecast multivariate series by learning channel dependence with graph neural networks.
87 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Long-term forecasting, keyword cluster “graph · multivariate time · spatial”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
22 papers in 2023–24 and 52 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 21.4% to 21.3%.
Open questions the data points to
- Papers report PatchTST on ETTm2 (MAE) anywhere from 0.145 to 0.326: comparisons against it depend on whose number you take.
- Papers report iTransformer on Electricity (MAE) anywhere from 0.212 to 0.27: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Learning Adaptive Forecasting Models from Irregularly Sampled Multivariate Clinical Data (AAAI 2016 · 30 citations)
- 2018 · Hierarchical Deep Generative Models for Multi-Rate Multivariate Time Series (ICML 2018 · 26 citations)
- 2020 · DATA-GRU: Dual-Attention Time-Aware Gated Recurrent Unit for Irregular Multivariate Time Series (AAAI 2020 · 123 citations)
- 2021 · Discrete Graph Structure Learning for Forecasting Multiple Time Series (ICLR 2021 · 98 citations)
- 2022 · CATN: Cross Attentive Tree-Aware Network for Multivariate Time Series Forecasting (AAAI 2022 · 47 citations)
- 2023 · FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective (NeurIPS 2023 · 61 citations)
- 2024 · MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · 234 citations)
- 2025 · FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting (AAAI 2025 · 22 citations)
- 2026 · FreDN: Spectral Disentanglement for Time Series Forecasting via Learnable Frequency Decomposition (AAAI 2026 · 7 citations)
Earliest papers here
- Learning Adaptive Forecasting Models from Irregularly Sampled Multivariate Clinical Data (AAAI 2016 · 30 citations)
- Hierarchical Deep Generative Models for Multi-Rate Multivariate Time Series (ICML 2018 · 26 citations)
- Directional Label Rectification in Adaptive Graph (AAAI 2018 · 0 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
Electricity · MAE (lower is better)
4 papers, 2020–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 0.04 | Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting (NeurIPS 2020 · 116 citations) |
What has been tried
43 papers here report 394 measured changes (a component added, removed or swapped, compared with a baseline): helped 73%, no clear effect 5%, hurt 17%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add adaptive graph convolution (graph learning) | Flight, Weather, and ETTm2 (average over prediction lengths 96, 336) · MSE | 0.302 -> 0.195; 0.232 -> 0.218; 0.253 -> 0.245 | Helped | MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · 234 citations) |
| Use FC graph construction and graph convolution rather than separate per-patch graphs and convolutions | C-MAPSS FD001 · RMSE | 12.58±0.22 -> 11.62±0.19 | Helped | Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data (AAAI 2024 · 114 citations) |
| Add STAR channel interaction to a channel-independent MLP | Traffic · MSE | 0.442 -> 0.409 | Helped | SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion (NeurIPS 2024 · 89 citations) |
| add Static Global Filtering | ETTm1, Weather, Electricity · MSE | 0.392 -> 0.385; 0.250 -> 0.244; 0.197 -> 0.180 | Helped | FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting (AAAI 2025 · 22 citations) |
| add the Local Extractor | UEA Setting 1, 10 datasets · average accuracy | 0.737 -> 0.754 | Helped | MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · 20 citations) |
| add HTPGM and use the combined prediction and reconstruction objective | ETTh2, prediction length 96 · MAE | 0.368 -> 0.338 | Helped | Considering Nonstationary within Multivariate Time Series with Variational Hierarchical Transformer for Forecasting (AAAI 2024 · 10 citations) |
| add cross-domain representation learning | KDD 10% missing · RMSE | 0.4210 -> 0.4057 | Helped | Frequency-aware Generative Models for Multivariate Time Series Imputation (NeurIPS 2024 · 10 citations) |
| Replace MSS with a linear Query/Key projection | Long-term MTSF benchmarks · MSE | Hurt | Revisiting Attention for Multivariate Time Series Forecasting (AAAI 2025 · 9 citations) | |
| Use local channel correlation in embedding learning instead of removing it | P19 · AUROC | 88.0 ± 3.1 -> 89.5 ± 1.9 | Helped | TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · 9 citations) |
| add autoregressive self-attention | ETTm1, Electricity, Traffic, Weather, Solar-Energy (F=96) · MSE | 0.337 -> 0.320; 0.148 -> 0.139; 0.438 -> 0.424; 0.168 -> 0.158; 0.211 -> 0.202 | Helped | Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling (ICML 2024 · 7 citations) |
| Remove variable-modality discrimination loss (LDis) | UEA classification average · accuracy | 71.9 -> 70.1 | Hurt | Addressing Spatial-Temporal Heterogeneity: General Mixed Time Series Analysis via Latent Continuity Recovery and Alignment (NeurIPS 2024 · 7 citations) |
| Add Channel MSA and its adapter | ETTh1 · MSE | 0.431 -> 0.429 | Helped | CSformer: Combining Channel Independence and Mixing for Robust Multivariate Time Series Forecasting (AAAI 2025 · 6 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| PatchTST | ETTm2 | MAE | 0.145 – 0.326 | 2 papers, e.g. NeurIPS 2024 and AAAI 2025 |
| iTransformer | Electricity | MAE | 0.212 – 0.27 | 2 papers, e.g. NeurIPS 2024 and AAAI 2025 |
| GraFITi | MIMIC-III | MSE | 0.396 – 0.453 | 2 papers, e.g. AAAI 2025 and ICML 2025 |
| GraFITi | MIMIC-IV | MSE | 0.225 – 0.245 | 2 papers, e.g. AAAI 2025 and ICML 2025 |
| GraFITi | Physionet12 | MSE | 0.286 – 0.306 | 2 papers, e.g. AAAI 2025 and ICML 2025 |
| PatchTST | ETTm1 | MAE | 0.379 – 0.4 | 2 papers, e.g. NeurIPS 2024 and AAAI 2025 |
What reviewers will expect
Datasets and baselines named in the 19 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Weather · 7 of 19 papers
- ETTh1 · 6 of 19 papers
- Traffic · 6 of 19 papers
- ETTh2 · 5 of 19 papers
- ETTm1 · 5 of 19 papers
- ETTm2 · 5 of 19 papers
- Electricity · 5 of 19 papers
- Solar-Energy · 4 of 19 papers
Baselines
- iTransformer · 8 of 19 papers
- PatchTST · 8 of 19 papers
- Crossformer · 7 of 19 papers
- TimesNet · 7 of 19 papers
- DLinear · 5 of 19 papers
- SCINet · 4 of 19 papers
- FEDformer · 4 of 19 papers
- Autoformer · 3 of 19 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 74 papers; from OpenAlex and the papers' affiliations).
- Harbin Institute of Technology · 5 papers
- Shanghai Jiao Tong University · 3 papers
- University of Electronic Science and Technology of China · 3 papers
- Nanyang Technological University · 3 papers
- Beijing Institute of Technology · 3 papers
- Nankai University · 3 papers
- University of Science and Technology of China · 3 papers
- South China University of Technology · 3 papers
- The Chinese University of Hong Kong · 2 papers
- Sichuan University · 2 papers
Use it
Download BibTeX of 33 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Attention for time series forecasting (123 papers)
- Spatiotemporal forecasting models (77 papers)
Elsewhere in the topic
- Lightweight long-term forecasters (94 papers)
- Transformers for long-term forecasting (25 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).