Research map / Graph learning / Knowledge and temporal graphs
Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph 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 EAC, a continual spatio-temporal graph forecasting method that freezes an STGNN backbone and tunes a node-level prompt parameter pool grown by expansion and compressed via low-rank factorization.
Problem
Continual forecasting on streaming spatio-temporal graphs whose node sets expand over time, where fine-tuning STGNNs causes catastrophic forgetting and high tuning/storage cost.
Approach
Freeze the STGNN backbone and learn only a node-level prompt parameter pool that grows with newly added nodes (expand, motivated by a heterogeneity/dispersion argument) and is factorized into low-rank node subspace matrices plus a shared matrix (compress, motivated by a low-rank argument with k=O(log min(n,d))).
Key results
- On PEMS-Stream average, EAC reaches MAE 13.53 and RMSE 21.77 versus 14.14/23.04 for the best prior baseline STKEC.
- EAC attains lowest MAE on all three datasets, e.g., Air-Stream average MAE 20.75 vs 21.12 for Continual-ST-AN.
- Training time per period is reduced by a factor of 1.26 to 3.02 relative to other methods, and EAC uses about 59% of tuning parameters at k=6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the EAC continuous prompt pool instead of continual full-model tuning | Continual spatio-temporal forecasting · Air-Stream, PEMS-Stream, and Energy-Stream | MAE, RMSE, and MAPE: Air-Stream Avg. MAE: 21.12 ±0.51 -> 20.75 ±0.20; PEMS-Stream Avg. MAE: 14.14 ±0.04 -> 13.53 ±0.06; Energy-Stream Avg. MA | Helped |
| Apply EAC prompt pooling rather than LoRA layers | PEMS-Stream forecasting with the same backbone · PEMS-Stream 12-step | MAE: 16.22 ±0.13 -> 14.92 ±0.11 | Helped |
| Apply EAC prompt pooling rather than LoRA layers | PEMS-Stream forecasting with the same backbone · PEMS-Stream 12-step | Avg. Time: 337.31 ±23.90 -> 224.33 ±26.35 | Helped |
| Add EAC prompts to different STGNN backbones | Six spatial/spectral graph and sequence operator combinations · PEMS-Stream Avg. | MAE: 14.07 ±0.07 -> 13.72 ±0.01; 13.23 ±0.10 -> 12.83 ±0.05; 14.69 ±0.42 -> 14.64 ±0.53; 14.01 ±0.01 -> 13.62 ±0.12; 13.73 ±0 | Mixed |
| Reduce the prompt-pool rank k from 6 to 2 (ablation) | EAC · Larger datasets | Tuning parameters: ∼ 63% | Helped |
| Decrease prompt-pool rank k (ablation) | EAC on PEMS-Stream · PEMS-Stream | Forecasting performance | Hurt |
| Increase prompt-pool rank k beyond 6 (ablation) | EAC on PEMS-Stream · PEMS-Stream | Forecasting performance | No clear effect |
Surprising findings
- Pretrain-ST performs worst on small datasets, and Continual-ST-NN (new nodes only) performs poorly due to forgetting.
- Attention-based backbones benefit least from EAC, attributed to overfitting from extra parameters.
Limitations
- Some baselines (PECPM, TFMoE, STKEC) lacked official code, so comparisons may be unfair or reproduced by the authors.
- Only one hyper-parameter k is studied, and the theoretical low-rank result is an existence argument rather than a guarantee on trained prompts.
- Few-shot and universality results are limited to PEMS-Stream in the main text.
Experimental setup
Datasets: PEMS-Stream Air-Stream Energy-Stream
Benchmarks: PEMS-Stream Air-Stream Energy-Stream
Metrics: MAE RMSE MAPE
Baselines: Pretrain-ST Retrain-ST Continual-ST-AN Continual-ST-NN TrafficStream STKEC PECPM TFMoE LoRA-based fine-tuning
Builds on: TrafficStream TFMoE STKEC PECPM LoRA
Closest papers by meaning
- Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling (ICML 2024 · similarity 0.89)
- ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data (AAAI 2025 · similarity 0.88)
- How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning (NeurIPS 2025 · similarity 0.87)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.87)
- Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting (NeurIPS 2025 · similarity 0.87)
- Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting (ICML 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.