Research map / Graph learning / GNN architectures
Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction
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 a dynamics-informed GNN (FA-GCN) that uses variant prevalence-ratio time series and global growth rates as node features to predict when a COVID-19 variant will reach a dominance threshold in a country before it arrives, and releases a benchmarking pipeline.
Problem
Predicting when a COVID-19 variant already circulating elsewhere will arrive and reach a dominance threshold in a region that has zero prevalence of it so far.
Approach
Derives variant prevalence-ratio dynamics (log prevalence ratio grows linearly in time) and uses them as node features (log prevalence-ratio time series plus global growth rate) in a GNN over a mobility graph with border-control edge weights, rather than as a physics loss regularizer as in PINNs.
Key results
- FA-GCN achieved the lowest MedMedAE (1.27) and MedMAE (1.29) among all compared models, versus GCN at 1.32/1.43 and the Dynamics-based Model at 1.86/2.9.
- PI-GCN (physics-informed loss) scored 2.24 MedMedAE and 2.01 MedMAE, worse than the plain GCN, indicating dynamics-as-regularization underperforms dynamics-as-features.
- Ablation shows including the global growth rate S as a feature improves results across most models; [r|S] combination performed best among feature sets except for Decision Tree.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use dynamics-informed feature augmentation in place of a standard GCN | Delay Prediction · Delay Prediction | MedMedAE: 1.32 -> 1.27 | Helped |
| use dynamics-informed feature augmentation rather than PINN-style loss regularization | Delay Prediction · Delay Prediction | MedMAE: 2.01 -> 1.29 | Helped |
| use dynamics-informed feature augmentation rather than the dynamics-based model | Delay Prediction · Delay Prediction | MedMedAE: 1.86 -> 1.27 | Helped |
| use dynamics-informed feature augmentation rather than a GRU | Delay Prediction · Delay Prediction | MedMedAE: 1.5 -> 1.27 | Helped |
| add global growth rate S to prevalence-ratio time-series features r (ablation) | Decision Tree · Delay Prediction | MedMedAE: 3.0 -> 3.48 | Hurt |
| add global growth rate S to prevalence time-series features p (ablation) | Decision Tree · Delay Prediction | MedMedAE: 3.0 -> 2.81 | Helped |
| add global growth rate S to prevalence-ratio time-series features r (ablation) | MLP · Delay Prediction | MedMedAE: 2.67 -> 2.38 | Helped |
| add global growth rate S to prevalence time-series features p (ablation) | MLP · Delay Prediction | MedMedAE: 2.55 -> 2.69 | Hurt |
| add global growth rate S to prevalence-ratio time-series features r (ablation) | GRU · Delay Prediction | MedMedAE: 1.50 -> 1.50 | No clear effect |
| use prevalence-ratio time-series features r instead of prevalence time-series features p, with S (ablation) | GRU · Delay Prediction | MedMedAE: 1.95 -> 1.50 | Helped |
| add global growth rate S to prevalence-ratio time-series features r (ablation) | GCN · Delay Prediction | MedMedAE: 1.50 -> 1.32 | Helped |
| use prevalence-ratio time-series features r instead of prevalence time-series features p, with S (ablation) | GCN · Delay Prediction | MedMedAE: 1.78 -> 1.32 | Helped |
Surprising findings
- The physics-informed loss variant (PI-GCN) did not beat its plain GCN counterpart, suggesting dynamics are more useful as input features than as loss regularization.
- The dynamics-based linear baseline had a MedMAE (2.9) much worse than its MedMedAE (1.86), indicating sensitivity to outlier predictions.
Limitations
- Only predicts delay conditional on eventual dominance; does not handle variants that never become dominant in a country (Dominance Prediction left to future work).
- Some variants (e.g., 21J.Delta, 23E.Omicron) lingering in few countries may reflect data-capture issues, a major source of error.
- Border-control data (OxCGRT) extended beyond Jan 2023 by assumption that restrictions stayed unchanged.
Experimental setup
Datasets: CoVariants OxCGRT
Benchmarks: CoVariants delay prediction benchmark
Metrics: MedMedAE MedMAE
Baselines: Mean Model Dynamics-based Model Decision Tree MLP GRU GCN PI-GCN
Builds on: Physics-informed neural networks Graph Convolutional Network GraphNorm GRU
Closest papers by meaning
- EARTH: Epidemiology-Aware Neural ODE with Continuous Disease Transmission Graph (ICML 2025 · similarity 0.84)
- SNN-PDE: Learning Dynamic PDEs from Data with Simplicial Neural Networks (AAAI 2024 · similarity 0.82)
- Graph Mixture of Experts and Memory-augmented Routers for Multivariate Time Series Anomaly Detection (AAAI 2025 · similarity 0.81)
- Valid Conformal Prediction for Dynamic GNNs (ICLR 2025 · similarity 0.81)
- CEGA: A Cost-Effective Approach for Graph-Based Model Extraction and Acquisition (ICML 2025 · similarity 0.81)
- When Graph Neural Networks Meet Dynamic Mode Decomposition (ICLR 2025 · similarity 0.81)
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.