Research map / Graph learning / Knowledge and temporal graphs
Multi-Scale Graph Learning for Anti-Sparse Downscaling
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 Multi-Scale Graph Learning (MSGL) with coarse, cross-scale interpolation, and fine-scale tasks plus an asynchronous pretrain-on-coarse-predictions variant (ASYNC-MSGL) for anti-sparse stream temperature downscaling in the Delaware River Basin.
Problem
Fine-scale (≤1 km) daily stream temperature prediction is hard because fine-scale observations are extremely sparse, and existing multi-resolution methods are designed for gridded data rather than graph-structured stream networks.
Approach
A multi-task graph model (shared RGrN embedding with coarse, cross-scale interpolation via a cross-scale distance matrix, and fine-scale heads) trained with MGDA-based gradient weighting, plus an asynchronous scheme that pretrains on coarse-model predictions remapped to fine reaches before fine-tuning on sparse observations.
Key results
- At 1% fine-scale labels, ASYNC-MSGL achieved RMSE 1.360 on Lower Delaware, 2.179 on LW Branch Delaware, 1.617 on Neversink, and 1.482 on Rancocas, outperforming RGrN and multi-scale baselines MS-STGCN and MSGNET in most watersheds.
- At 0.1% labels in Lower Delaware, ASYNC-MSGL pretrained on D_mapping reduced RMSE from 2.521 to 1.818 relative to MSGL without pretraining.
- Ablation shows removing CSL (MSGL w/o CSL) or CrSL (w/o CrSL) degrades performance versus full MSGL, and MSO helps in most watersheds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cross-scale interpolation learning (CrSL) to CSL and FSL (ablation) | MSGL · Delaware watersheds | RMSE | Helped |
| Add coarse-scale graph learning (CSL) to CrSL and FSL (ablation) | MSGL · Delaware watersheds | RMSE | Helped |
| Add multi-scale optimization (MSO) (ablation) | MSGL · Lower West Branch Delaware, Neversink, Rancocas, Lower Delaware | RMSE | Helped |
| Use asynchronous mapped-prediction pretraining instead of synchronous MSGL | MSGL · Lower Delaware, LwBranch Delaware, Neversink, Rancocas (1% labels) | RMSE: 1.384 -> 1.360; 2.375 -> 2.179; 1.919 -> 1.617; 1.450 -> 1.482 | Mixed |
| Pretrain on mapped coarse predictions instead of no pretraining | MSGL · Lower Delaware, LwBranch Delaware, Neversink, Rancocas (0.1% labels) | RMSE: 2.521 -> 1.818; 2.687 -> 2.376; 2.323 -> 2.036; 4.431 -> 2.794 | Helped |
| Pretrain on mapped coarse predictions instead of no pretraining | MSGL · Lower Delaware, LwBranch Delaware, Neversink, Rancocas (1% labels) | RMSE: 1.384 -> 1.360; 2.375 -> 2.179; 1.919 -> 1.617; 1.450 -> 1.483 | Mixed |
| Pretrain on Brandywine-Christina data | MSGL · Four watersheds (0.1% and 1% labels) | RMSE: 2.521 -> 2.006; 1.384 -> 1.408; 2.687 -> 2.936; 2.375 -> 2.495; 2.323 -> 2.664; 1.919 -> 1.759; 4.431 -> 3.855; 1.450 -> | Mixed |
Surprising findings
- Cross-domain pretraining on Brandywine-Christina gave inconsistent gains and sometimes degraded performance, while D_mapping asynchronous pretraining was more consistent.
- Removing MSO (w/o MSO) gave the best RMSE for Lower Delaware at 1% labels (1.391), slightly better than full MSGL (1.384 is marginally better; reported bolded tie).
Limitations
- Evaluated only on five well-observed Delaware River Basin watersheds, with sparsity simulated by masking labels.
- Asynchronous pretraining depends on a coarse-scale regional model and on a remapping assumption between coarse and fine reaches.
- Gains from cross-domain pretraining are inconsistent; ASYNC-MSGL loses to MSGL at 1% on Rancocas.
Experimental setup
Datasets: Delaware River Basin stream temperature observations (Lower Delaware, LW Branch Delaware, Neversink, Rancocas, Brandywin NHM geospatial fabric (coarse segments) NHD (fine segments)
Benchmarks: Delaware River Basin watersheds (Lower Delaware, LW Branch Delaware, Neversink, Rancocas)
Metrics: RMSE standard deviation over nine replicates Welch's t-test
Baselines: RGrN STGCN DCRNN GASTN MS-STGCN MSGNET
Builds on: RGrN MGDA
Closest papers by meaning
- Fair Graph Learning Using Constraint-Aware Priority Adjustment and Graph Masking in River Networks (AAAI 2024 · similarity 0.88)
- Physics-Guided Fair Graph Sampling for Water Temperature Prediction in River Networks (AAAI 2025 · similarity 0.88)
- The Merit of River Network Topology for Neural Flood Forecasting (ICML 2024 · similarity 0.84)
- DroughtSet: Understanding Drought Through Spatial-Temporal Learning (AAAI 2025 · similarity 0.83)
- Topology-Aware Conformal Prediction for Stream Networks (NeurIPS 2025 · similarity 0.83)
- Learning from Polar Representation: An Extreme-Adaptive Model for Long-Term Time Series Forecasting (AAAI 2024 · similarity 0.83)
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.