Research map / Time series / Long-term forecasting
TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for 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 TimeFilter, a GNN-based forecaster that builds a patch-level spatial-temporal graph and uses a patch-specific MoE-style dynamic router to filter dependencies, reporting SOTA on 13 datasets.
Problem
Multivariate time series forecasting must balance channel-independent and channel-dependent modeling; existing channel-clustering methods are coarse-grained and cannot capture time-varying inter-channel dependencies.
Approach
Construct a k-NN spatial-temporal graph over channel-patch tokens, decompose it into patch-specific ego graphs split into temporal, spatial and spatial-temporal regions, and use a dynamic-allocation MoE router (top-p cumulative confidence) to keep only the relevant dependency types per patch before GNN aggregation.
Key results
- Long-term forecasting with input length 96: TimeFilter reduces MSE by 4.48% and MAE by 2.23% relative to the second-best model Leddam, averaged across datasets.
- Short-term PEMS forecasting: on PEMS08 TimeFilter reduces MSE and MAE by 13.54% and 3.13% versus the second-best DUET.
- Under searched look-back lengths, TimeFilter reduces MSE/MAE by 5.34%/1.40% versus DUET and 6.89%/3.69% versus CCM (as reported by the authors).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use patch-specific spatial-temporal filtration | TimeFilter · Long-term forecasting, L=96, ETT (Avg.), Weather, Electricity, Traffic, Solar-Energy | MSE: ETT (Avg.): 0.358 vs 0.368; Weather: 0.239 vs 0.242; Electricity: 0.158 vs 0.169; Traffic: 0.407 vs 0.451; Solar-Energy: | Helped |
| Use TimeFilter rather than the strongest listed baseline | TimeFilter · Long-term forecasting with searched look-back length | MSE: Weather: 0.216 vs 0.218; Electricity: 0.150 vs 0.158; Traffic: 0.360 vs 0.394; Climate: 1.047 vs 1.060 | Helped |
| Use TimeFilter rather than the strongest listed baseline | TimeFilter · PEMS03, PEMS04, PEMS07, PEMS08 | MSE: PEMS03: 0.084 vs 0.086; PEMS04: 0.083 vs 0.085; PEMS07: 0.071 vs 0.075; PEMS08: 0.083 vs 0.096 | Helped |
| Replace patch-specific filtration with Top-K edge selection (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.243 | Hurt |
| Replace patch-specific filtration with random-K edge selection (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.243 | Hurt |
| Replace patch-specific filtration with region-wise Top-K selection (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.241 | Hurt |
| Replace patch-specific filtration with learned region thresholds (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.241 | Hurt |
| Replace patch-specific filtering with channel-wise filtering (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.242 | Hurt |
| Remove filtration after graph construction (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.244 | Hurt |
| Replace dynamic routing with static routing selecting only the most significant filter (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.242 | Hurt |
| Remove importance and dynamic routing losses (ablation) | TimeFilter · Weather | MSE: 0.239 -> 0.243 | Hurt |
Surprising findings
- Higher-weight graph edges may reflect spurious regression rather than effective dependencies, so magnitude-based filtering (Top-K) underperforms the learned patch-specific filter.
Limitations
- Hyperparameters such as the scaling factor alpha and Top-p are hand-tuned or selected from small grids.
- Ablations and error bars are reported on a subset of datasets and three runs only.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic Solar-Energy Climate PEMS03 PEMS04 PEMS07
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic Solar-Energy Climate PEMS03 PEMS04 PEMS07
Metrics: MSE MAE
Baselines: DUET iTransformer Leddam MSGNet SOFTS CrossGNN PatchTST Crossformer TimesNet DLinear FEDformer CCM
Builds on: Mixture of Experts PatchTST CCM DUET MoE-Dynamic
Closest papers by meaning
- From Similarity to Superiority: Channel Clustering for Time Series Forecasting (NeurIPS 2024 · similarity 0.89)
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.88)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.88)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.87)
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.86)
- TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting (NeurIPS 2026 · 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.