Research map / Graph learning / GNN architectures
GRAIN: Multi-Granular and Implicit Information Aggregation Graph Neural Network for Heterophilous Graphs
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 GRAIN, a GCN-based heterophily-aware node classifier that uses TD3 reinforcement learning to pick continuous aggregation hop counts per node and adds implicit distant-node aggregation, evaluated on 13 homophilous and heterophilous datasets.
Problem
GNNs underperform on heterophilous graphs where the homophily assumption fails; existing methods overlook aggregation at different information granularities and implicit relationships between distant non-neighboring nodes.
Approach
Models node aggregation hop count as a continuous action in an MDP solved by TD3 actor-critic (reward = windowed validation accuracy), and aggregates multi-hop granularity plus implicit distant-node information via a multi-view aggregator with parameter sharing and a buffer for efficiency.
Key results
- GRAIN reaches 88.52% on Cora, 81.25% on Citeseer, and 90.12% on Cornell, the best among the compared methods on those datasets.
- Reported overall improvement of 2.36% over the best baseline (GNN-SATA) on the eight-dataset table.
- GRAIN achieves the best accuracy on six homophilous datasets in the appendix table, e.g., 87.16% on Computers and 95.59% on Coauthor CS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRAIN instead of the strongest baseline | GCN node classification · Pubmed | accuracy: 86.91 -> 87.04 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Citeseer | accuracy: 79.81 -> 81.25 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Cora | accuracy: 86.72 -> 88.52 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Cornell | accuracy: 88.89 -> 90.12 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Actor | accuracy: 39.66 -> 38.89 | Hurt |
| Use GRAIN instead of the strongest baseline | GCN node classification · Wisconsin | accuracy: 88.54 -> 89.01 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Texas | accuracy: 87.69 -> 87.69 | No clear effect |
| Use GRAIN instead of the strongest baseline | GCN node classification · Chameleon | accuracy: 56.12 -> 56.43 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Film | accuracy: 39.33 -> 39.77 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Squirrel | accuracy: 38.75 -> 38.96 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Computers | accuracy: 86.98 -> 87.16 | Helped |
| Use GRAIN instead of the strongest baseline | GCN node classification · Photo | accuracy: 93.66 -> 93.79 | Helped |
Surprising findings
- MLP outperforms several GNNs on highly heterophilous datasets such as Cornell, Wisconsin and Texas.
- GNN-SATA runs out of memory on Pubmed, Computers and Coauthor CS.
Limitations
- Iterative training to learn per-node granularity is resource-intensive and scaling to much larger graphs remains challenging.
- The next state is sampled randomly, so no optimality guarantee is provided.
- Only a basic GCN backbone is used to isolate the aggregator's effect, so generality to other architectures is untested.
Experimental setup
Datasets: Cora Citeseer Pubmed Computers Photo Coauthor CS Cornell Wisconsin Texas Film Actor Squirrel
Benchmarks: Cora Citeseer Pubmed Computers Photo Coauthor CS Cornell Wisconsin Texas Film Actor Squirrel
Metrics: node classification test accuracy average accuracy drop
Baselines: MLP GCN GAT GraphSAGE MixHop APPNP GCNII GPR-GNN GGCN ACM-GNN FE-GNN GNN-SATA
Builds on: GCN TD3 GPR-GNN ACM-GNN GCNII
Closest papers by meaning
- Integrating Co-Training with Edge Discrimination to Enhance Graph Neural Networks Under Heterophily (AAAI 2025 · similarity 0.88)
- Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.88)
- Improving Graph Neural Networks by Learning Continuous Edge Directions (ICLR 2025 · similarity 0.88)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.88)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · similarity 0.88)
- On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.88)
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.