Heterophily Benchmark
Amazon-ratings
Structure
1Graphs
24,492Nodes
93,050Edges
5Classes
300Node Feature Dim
0.38Homophily
Coverage
259Results
1Variants
35Papers
- Graph Cascades: Contagion-Based Mesoscopic Rewiring for Structure-Aware Graph Machine Learning Graph Cascades: Contagion-Based Mesoscopic Rewiring for Structure-Aware Graph Machine Learning
- Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification
- Geometric Evolution Graph Convolutional Networks: Enhancing Graph Representation Learning via Ricci Flow Geometric Evolution Graph Convolutional Networks: Enhancing Graph Representation Learning via Ricci Flow
- Torch Geometric Pool: the Pytorch library for pooling in Graph Neural Networks Torch Geometric Pool: the Pytorch library for pooling in Graph Neural Networks
- ICML 2025: View Space: Learning Representation across Arbitrary Graphs View Space: Learning Representation across Arbitrary Graphs ICML 2025
- NeurIPS 2025: GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks NeurIPS 2025
- NeurIPS 2025: Relieving the Over-Aggregating Effect in Graph Transformers Relieving the Over-Aggregating Effect in Graph Transformers NeurIPS 2025
- Gumbel-MPNN: Graph Rewiring with Gumbel-Softmax Gumbel-MPNN: Graph Rewiring with Gumbel-Softmax
- Aligning the Spectrum: Hybrid Graph Pre-training and Prompt Tuning across Homophily and Heterophily Aligning the Spectrum: Hybrid Graph Pre-training and Prompt Tuning across Homophily and Heterophily
- TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows
- KDD 2025: Discrepancy-Aware Graph Mask Auto-Encoder Discrepancy-Aware Graph Mask Auto-Encoder KDD 2025
- TMLR 2025: Piecewise Constant Spectral Graph Neural Network Piecewise Constant Spectral Graph Neural Network TMLR 2025
- TMLR 2025: Heterophily-informed Message Passing Heterophily-informed Message Passing TMLR 2025
- ICML 2025: Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks ICML 2025
- NeurIPS 2025: A Signed Graph Approach to Understanding and Mitigating Oversmoothing in GNNs A Signed Graph Approach to Understanding and Mitigating Oversmoothing in GNNs NeurIPS 2025
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning
- NeurIPS 2024: Even Sparser Graph Transformers Even Sparser Graph Transformers NeurIPS 2024
- NeurIPS 2024: Distributed-Order Fractional Graph Operating Network Distributed-Order Fractional Graph Operating Network NeurIPS 2024
- ICML 2024: Diss-l-ECT: Dissecting Graph Data with local Euler Characteristic Transforms Diss-l-ECT: Dissecting Graph Data with local Euler Characteristic Transforms ICML 2024
- Re-evaluating the Advancements of Heterophilic Graph Learning Re-evaluating the Advancements of Heterophilic Graph Learning
- UniGAP: A Universal and Adaptive Graph Upsampling Approach to Mitigate Over-Smoothing in Node Classification Tasks UniGAP: A Universal and Adaptive Graph Upsampling Approach to Mitigate Over-Smoothing in Node Classification Tasks
- Scalable Graph Compressed Convolutions Scalable Graph Compressed Convolutions
- ICLR 2024: Revisiting Random Walks for Learning on Graphs Revisiting Random Walks for Learning on Graphs ICLR 2024
- ICML 2024: Sign is Not a Remedy: Multiset-to-Multiset Message Passing for Learning on Heterophilic Graphs Sign is Not a Remedy: Multiset-to-Multiset Message Passing for Learning on Heterophilic Graphs ICML 2024
- Differentiable Cluster Graph Neural Network Differentiable Cluster Graph Neural Network
- Bundle Neural Networks for message diffusion on graphs Bundle Neural Networks for message diffusion on graphs
- ICLR 2024: Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND ICLR 2024
- TMLR 2024: Contextualized Messages Boost Graph Representations Contextualized Messages Boost Graph Representations TMLR 2024
- ICLR 2024: Polynormer: Polynomial-Expressive Graph Transformer in Linear Time Polynormer: Polynomial-Expressive Graph Transformer in Linear Time ICLR 2024
- KDD 2024: Graph Mamba: Towards Learning on Graphs with State Space Models Graph Mamba: Towards Learning on Graphs with State Space Models KDD 2024
- ICML 2024: Understanding Heterophily for Graph Neural Networks Understanding Heterophily for Graph Neural Networks ICML 2024
- ICML 2023: Evolving Computation Graphs Evolving Computation Graphs ICML 2023
- IJCAI 2023: Graph Neural Convection-Diffusion with Heterophily Graph Neural Convection-Diffusion with Heterophily IJCAI 2023
- Graph Neural Networks for Graphs with Heterophily: A Survey Graph Neural Networks for Graphs with Heterophily: A Survey
- Graph Prototypical Networks for Few-shot Learning on Attributed Networks Graph Prototypical Networks for Few-shot Learning on Attributed Networks
5-class node classification Amazon product co-purchase node classification (5 rating classes) from the heterophilic benchmark. Uses the corrected benchmark features and ten official splits. Evaluated by Accuracy.
Dataset details
Task
Amazon product co-purchase node classification (5 rating classes) from the heterophilic benchmark. The primary catalog metric is Accuracy.
Data
Amazon product co-purchase node classification (5 rating classes) from the heterophilic benchmark.
Features
The benchmark provides one fixed graph, node features, labels, and ten official splits per dataset. Features are dataset-specific and include text-derived, product, synthetic-grid, or platform-account attributes described by the release.
Splits and evaluation
The catalog records these protocols or variants: Standard split. Evaluation uses Accuracy (higher is better). Exact masks or folds must come from the cited release.
Quirks and pitfalls
Use the corrected release and its ten masks. Several tasks are imbalanced and use ROC-AUC rather than accuracy; treating missing relationships as negative evidence or tuning on the test masks invalidates comparison.
Sources
SOTA Progression
new SOTA
matches SOTA at that time
other proposed result
other baseline result
Leaderboard
| # | Model | Accuracy ↑ | ACC | AUROC | macro-F1 | Paper |
|---|---|---|---|---|---|---|
| 1 | UniGAP | 0.6375
±
0.0110 | — | — | — | |
| 2 | H2GCN + HalfHop | 0.5961
±
0.0100 | — | — | — | |
| 3 | GraphSAGE + HalfHop | 0.5897
±
0.0150 | — | — | — | |
| 4 | GraphSAGE + AdaEdge | 0.5884
±
0.0130 | — | — | — | |
| 5 | H2GCN + AdaEdge | 0.5877
±
0.0120 | — | — | — | |
| 6 | GCN +AdaEdge | 0.5692
±
0.0110 | — | — | — | |
| 7 | H2GCN | 0.5681
±
0.0090 | — | — | — | |
| 8 | SAGE+T | 0.5672
±
0.0035 | — | — | — | |
| 9 | GraphTARIF | 0.5586
±
0.0042 | — | — | — | |
| 10 | RWNN-Llama3-8b, fine-tuned | 0.5576 | — | — | — | |
| 11 | GAT+T | 0.5573
±
0.0031 | — | — | — | |
| 12 | SAGE | 0.5569
±
0.0023 | — | — | — | |
| 13 | ACMGCN | 0.5565
±
0.0176 | — | — | — | |
| 14 | GAT | 0.5559
±
0.0027 | — | — | — | |
| 15 | GraphSAGE | 0.5540
±
0.0020 | — | — | — | |
| 16 | GCN + HalfHop | 0.5519
±
0.0160 | — | — | — | |
| 17 | Polynormer + Wideformer | 0.5505
±
0.0008 | — | — | — | |
| 18 | Polynormer | 0.5496
±
0.0022 | — | — | — | |
| 19 | Polynormer +ENACT | 0.5494 | — | — | — | |
| 20 | Polynormer-r | 0.5481
±
0.0049 | — | — | — | |
| 21 | Polynormer | 0.5446
±
0.0040 | — | — | — | |
| 22 | Polynormer-v1 | 0.5446
±
0.0040 | — | — | — | |
| 23 | Polynormer-v2 | 0.5441
±
0.0035 | — | — | — | |
| 24 | SGFormer | 0.5414
±
0.0062 | — | — | — | |
| 25 | GMN | 0.5407
±
0.0031 | — | — | — | |
| 26 | GCN+TS | 0.5406
±
0.0058 | — | — | — | |
| 27 | Tango -GPS | 0.5383
±
0.0032 | — | — | — | |
| 28 | GCN | 0.5380
±
0.0060 | — | — | — | |
| 29 | BuNN | 0.5374
±
0.0051 | — | — | — | |
| 30 | ECG-GAT-sep | 0.5365
±
0.0039 | — | — | — | |
| 31 | MLP->GNN-ECG-GAT-sep | 0.5365
±
0.0039 | — | — | — | |
| 32 | GraphSAGE | 0.5363
±
0.0039 | — | — | — | |
| 33 | SAGE | 0.5363
±
0.0039 | — | — | — | |
| 34 | GCN-MAS | 0.5354
±
0.0159 AUROC
0.7710 macro-F1
0.4180 | — | 0.7710 | 0.4180 | |
| 35 | Exphormer | 0.5351
±
0.0046 | — | — | — | |
| 36 | Exphormer+GAT | 0.5351
±
0.0046 | — | — | — | |
| 37 | GraphSAGE+mix. | 0.5350
±
0.0070 | — | — | — | |
| 38 | SGFormer + Wideformer | 0.5347
±
0.0014 | — | — | — | |
| 39 | ECG-SAGE | 0.5345
±
0.0027 | — | — | — | |
| 40 | MLP->GNN-ECG-SAGE | 0.5345
±
0.0027 | — | — | — | |
| 41 | MLP-ECG-SAGE | 0.5332
±
0.0039 | — | — | — | |
| 42 | SGFormer +ENACT | 0.5332 | — | — | — | |
| 43 | ECG-GT-sep | 0.5325
±
0.0039 | — | — | — | |
| 44 | MLP->GNN-ECG-GT-sep | 0.5325
±
0.0039 | — | — | — | |
| 45 | AGS-GNN | 0.5321
±
0.0046 | — | — | — | |
| 46 | BGRL-ECG-SAGE | 0.5312
±
0.0032 | — | — | — | |
| 47 | GraphGPS | 0.5310
±
0.0042 | — | — | — | |
| 48 | GPS | 0.5310
±
0.0042 | — | — | — | |
| 49 | MLP-ECG-GAT-sep | 0.5298
±
0.0030 | — | — | — | |
| 50 | ACM-GCN | 0.5294
±
0.0023 | — | — | — |
SOTA performance
set a new SOTA at the time
matched the SOTA at the time