Research map / Graph learning / Knowledge and temporal graphs
Fair Graph Learning Using Constraint-Aware Priority Adjustment and Graph Masking in River Networks
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 Fair-Graph, an LSTM+GraphSAGE river-network model with learned aggregation/adaptation graph masks and a constraint-aware learning-rate priority adjustment to improve prediction and spatial fairness of water temperature and streamflow.
Problem
Graph-based data-driven models for water temperature and streamflow prediction in large river networks show spatial bias, with performance disparities across river segments due to heterogeneous data, sparse observations, and noisy sensors.
Approach
Combines learned node-specific aggregation and adaptation graph masks with a centralized referee that adjusts per-segment learning rates by performance deviation, capped by a per-segment performance upper-bound, so that training priority is not wasted on segments that cannot improve.
Key results
- On Delaware River Basin, Fair-Graph cpa+ma+mo achieves RMSE 1.792 vs 1.833 for Base, with fairness 0.537 vs 0.616; pa+ma+mo gets lowest RMSE 1.786.
- On Houston river network, Fair-Graph cpa+ma+mo reaches NSE 0.758 vs 0.721 for Base, though its fairness (0.087) is worse than the best variant pa (0.075).
- On perturbed DRB data, Fair-Graph cpa+ma+mo reduces RMSE to 1.953 from 2.037 for Base and fairness to 0.586 from 0.635.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add aggregation and adaptation masks (ablation) | Fair-Graph · Delaware River Basin | RMSE: 1.833 -> 1.795 | Helped |
| Add adaptation mask to aggregation-mask model (ablation) | Fair-Graph · Delaware River Basin | RMSE: 1.807 -> 1.795 | Helped |
| Use the adaptation mask instead of node-only fine-tuning for information selection (ablation) | Fair-Graph · Delaware River Basin | RMSE: 1.804 -> 1.795 | Helped |
| Add constraint-aware priority adjustment to graph masking and priority adjustment (ablation) | Fair-Graph · Houston River network | NSE: 0.738 -> 0.758 | Helped |
| Add constraint-aware priority adjustment to graph masking and priority adjustment (ablation) | Fair-Graph · Delaware River Basin | Fairness: 0.552 -> 0.537 | Hurt |
| Add graph masking to priority adjustment (ablation) | Fair-Graph · Delaware River Basin | RMSE: 1.830 -> 1.786 | Helped |
| Add graph masking to constraint-aware priority adjustment (ablation) | Fair-Graph · Houston River network | NSE: 0.737 -> 0.758 | Helped |
| Use priority adjustment without the performance upper-bound constraint (ablation) | Fair-Graph · Delaware River Basin | Fairness: 0.532 -> 0.556 | Helped |
| Use aggregation and adaptation masks with constraint-aware priority adjustment (ablation) | Fair-Graph · Delaware River Basin | RMSE: 1.833 -> 1.792 | Helped |
| Use aggregation and adaptation masks with constraint-aware priority adjustment (ablation) | Fair-Graph · Houston River network | NSE: 0.721 -> 0.758 | Helped |
| Use aggregation and adaptation masks with constraint-aware priority adjustment (ablation) | Fair-Graph · Delaware River Basin | Fairness: 0.616 -> 0.537 | Helped |
| Use aggregation and adaptation masks with constraint-aware priority adjustment (ablation) | Fair-Graph · Houston River network | Fairness: 0.113 -> 0.087 | Helped |
Surprising findings
- Adding the performance upper-bound constraint improved accuracy but reduced fairness relative to unconstrained priority adjustment (cpa vs pa), a trade-off the authors attribute to consistent training on already-saturated segments degrading others.
- Graph-masked variants improved accuracy without compromising fairness, with fairness roughly unchanged between priority-only and priority+mask variants.
Limitations
- Performance upper-bound is approximated from cross-validation with standard GNNs and is not a generalized estimate across models.
- Improvements in RMSE/NSE are small in absolute terms (e.g., Houston fairness improvements inconsistent across variants).
Experimental setup
Datasets: Delaware River Basin water temperature observations (USGS NWIS, Water Quality Portal) Houston River network streamflow from National Water Model v2.0 reanalysis NHDPlus river network
Benchmarks: Delaware River Basin water temperature (RMSE, fairness) Houston River network streamflow (NSE, fairness) perturbed DRB data with 90% observations dropped on 30 segments
Metrics: RMSE NSE spatial fairness measure (mean absolute deviation from overa
Baselines: Base LSTM+GNN REG ADL DSGNN
Builds on: GraphSAGE LSTM Fairness-aware GNN degree-specific parameters (DSGNN) spatial fairness partitioning (Xie et al. 2022)
Closest papers by meaning
- Physics-Guided Fair Graph Sampling for Water Temperature Prediction in River Networks (AAAI 2025 · similarity 0.91)
- Multi-Scale Graph Learning for Anti-Sparse Downscaling (AAAI 2025 · similarity 0.88)
- The Merit of River Network Topology for Neural Flood Forecasting (ICML 2024 · similarity 0.87)
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.85)
- Topology-Aware Conformal Prediction for Stream Networks (NeurIPS 2025 · similarity 0.84)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.84)
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.