Research map / Graph learning / Knowledge and temporal graphs

Improving Generalization of Dynamic Graph Learning via Environment Prompt

NeurIPS 2024 · Graph learning › Knowledge and temporal graphs · approach: Continuous-time dynamic graph learning · 4 citations (OpenAlex)

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 EpoD, a dynamic graph learning model that infers unseen environment variables via self-prompted learning and uses node-centered dynamic subgraphs as mediating variables in a structural causal model to improve temporal OOD generalization.

Problem

Temporal out-of-distribution generalization in dynamic graphs, where shifting unseen environments cause distribution shift; existing environment-based methods use predefined environment codebooks and only concatenate inferred environment features rather than modeling structural changes.

Approach

A self-prompted learning mechanism uses learnable node-wise prompt tokens and cross-attention to squeeze continuous environment representations from history without a predefined scale; these are then used to compute KL-based asymmetric node correlations that define node-centered L-hop dynamic subgraphs acting as the mediating variable in a spatio-temporal causal model with backdoor adjustment.

Key results

• On traffic 12→24 prediction, EpoD reports MAE 16.92 on PEMS08 and 21.12 on PEMS04, outperforming the listed baselines (best baseline CaST 17.28 and 20.79 on PEMS04 respectively).
• On social link prediction, EpoD reaches 80.85 AUC on Yelp and 83.85 on ACT, outperforming EAGLE (77.26 and 82.70); on COLLAB EAGLE is better (84.41 vs 83.21).
• Ablation: replacing dynamic subgraphs with no subgraph raises MAE on PEMS08 from 16.92 to 17.45 and drops Yelp AUC from 80.85 to 76.34.

Measured effects reported in the paper

Surprising findings

• On the COLLAB dataset, EAGLE (84.41 AUC) outperforms EpoD (83.21 AUC), which the paper does not emphasize.
• On PEMS04, EpoD's MAE (21.12) is nearly tied with CaST (20.79 in the table), where CaST is actually better in MAE per the table text, which the paper notes as optimal on PEMS04.

Limitations

• Strong dependence on graph topology: subgraph discovery is node filtering over K-hop neighbors, which lacks adaptability to changing relationships.
• Reliance on predefined L (set to 5) and beta (set to 0.2) hyperparameters, with sensitivity analysis only on selected datasets.
• Mixed results versus baselines: EAGLE outperforms EpoD on COLLAB; CaST is better on PEMS04 MAE in the table.

Experimental setup

Datasets: PEMS08 PEMS04 SD GBA COLLAB Yelp ACT EnvST

Benchmarks: PEMS08 PEMS04 SD GBA COLLAB Yelp ACT EnvST

Metrics: MAE RMSE AUC

Baselines: GWNET AGCRN Z-GCNETs DSTAGNN STGNCDE CaST DySAT IRM VREx GroupDRO DIDA EAGLE

Builds on: AGCRN DIDA EAGLE CaST prompt learning

Closest papers by meaning

• Enhancing Graph Invariant Learning from a Negative Inference Perspective (ICML 2025 · similarity 0.90)
• Expand Horizon: Graph Out-of-Distribution Generalization via Multi-Level Environment Inference (AAAI 2025 · similarity 0.87)
• Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.87)
• Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments (CVPR 2024 · similarity 0.87)
• Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.87)
• Node-Time Conditional Prompt Learning in Dynamic Graphs (ICLR 2025 · similarity 0.86)