OGB
ogbl-citation2
Structure
1Graphs
2,927,963Nodes
30,561,187Edges
128Node Feature Dim
Coverage
105Results
1Variants
20Papers
- Quantum Graph Attention Network: A Novel Quantum Multi-Head Attention Mechanism for Graph Learning Quantum Graph Attention Network: A Novel Quantum Multi-Head Attention Mechanism for Graph Learning
- Weak Models Can be Good Teachers: A Case Study on Link Prediction with MLPs Weak Models Can be Good Teachers: A Case Study on Link Prediction with MLPs
- NeurIPS 2024: On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks NeurIPS 2024
- Link Prediction with Untrained Message Passing Layers Link Prediction with Untrained Message Passing Layers
- KDD 2023: Efficient End-to-end Language Model Fine-tuning on Graphs Efficient End-to-end Language Model Fine-tuning on Graphs KDD 2023
- KDD 2023: LPFormer: An Adaptive Graph Transformer for Link Prediction LPFormer: An Adaptive Graph Transformer for Link Prediction KDD 2023
- Efficient Link Prediction via GNN Layers Induced by Negative Sampling Efficient Link Prediction via GNN Layers Induced by Negative Sampling
- Circle Feature Graphormer: Can Circle Features Stimulate Graph Transformer? Circle Feature Graphormer: Can Circle Features Stimulate Graph Transformer?
- Simplifying Distributed Neural Network Training on Massive Graphs: Randomized Partitions Improve Model Aggregation Simplifying Distributed Neural Network Training on Massive Graphs: Randomized Partitions Improve Model Aggregation
- JMLR 2023: Improving Graph Neural Networks on Multi-node Tasks with the Labeling Trick Improving Graph Neural Networks on Multi-node Tasks with the Labeling Trick JMLR 2023
- SUREL+: Moving from Walks to Sets for Scalable Subgraph-based Graph Representation Learning SUREL+: Moving from Walks to Sets for Scalable Subgraph-based Graph Representation Learning
- Asymmetric Learning for Graph Neural Network based Link Prediction Asymmetric Learning for Graph Neural Network based Link Prediction
- Simplifying Subgraph Representation Learning for Scalable Link Prediction Simplifying Subgraph Representation Learning for Scalable Link Prediction
- ICML 2022: Linkless Link Prediction via Relational Distillation Linkless Link Prediction via Relational Distillation ICML 2022
- ICLR 2022: Graph Neural Networks for Link Prediction with Subgraph Sketching Graph Neural Networks for Link Prediction with Subgraph Sketching ICLR 2022
- NeurIPS 2022: Neo-GNNs: Neighborhood Overlap-aware Graph Neural Networks for Link Prediction Neo-GNNs: Neighborhood Overlap-aware Graph Neural Networks for Link Prediction NeurIPS 2022
- Pairwise Learning for Neural Link Prediction Pairwise Learning for Neural Link Prediction
- Adaptive Graph Diffusion Networks Adaptive Graph Diffusion Networks
- NeurIPS 2020: Labeling Trick: A Theory of Using Graph Neural Networks for Multi-Node Representation Learning Labeling Trick: A Theory of Using Graph Neural Networks for Multi-Node Representation Learning NeurIPS 2020
- NeurIPS 2020: OGB OGB NeurIPS 2020
Link prediction For each 2019 source paper, rank a withheld true reference above 1,000 older, uncited candidate papers. Papers have 128-dimensional averages of skip-gram embeddings from their titles and abstracts. Evaluated by MRR.
Dataset details
Task
For each 2019 source paper, rank a withheld true reference above 1,000 older, uncited candidate papers. Mean reciprocal rank is averaged over the source-paper queries.
Data
The directed MAG subset contains 2,927,963 papers and 30,561,187 citation edges. Each node has its publication year and citations point from the citing paper to the referenced paper.
Features
Papers have 128-dimensional averages of skip-gram embeddings from their titles and abstracts. Candidate negatives are restricted to previous papers not already cited by the source.
Splits and evaluation
Two references are randomly removed from each eligible 2019 source paper, one for validation and one for test; all remaining citations train the model. Evaluation ranks each held-out target separately against its fixed negative candidates.
Quirks and pitfalls
The deprecated ogbl-citation dataset had faulty negative sampling and must not be combined with ogbl-citation2. Sampled MRR is not full-corpus retrieval, and sampled uncited papers are not necessarily irrelevant. Edge direction and temporal eligibility are essential to the protocol.
Sources
SOTA Progression
new SOTA
matches SOTA at that time
other proposed result
other baseline result
Leaderboard
| # | Model | MRR ↑ | Hits@200 | Paper |
|---|---|---|---|---|
| 1 | CFG | 0.8997
±
0.0015 | — | |
| 2 | SIEG | 0.8987
±
0.0018 | — | |
| 3 | LPFormer | 0.8981
±
0.0013 | — | |
| 4 | SUREL | 0.8974
±
0.0018 | — | |
| 5 | NGNN + SEAL | 0.8891
±
0.0022 | — | |
| 6 | SUREL | 0.8883
±
0.0018 | — | |
| 7 | PoS+ | 0.8814
±
0.0008 | — | |
| 8 | S3GRL (PoS Plus) | 0.8814
±
0.0008 | — | |
| 9 | SEAL | 0.8767
±
0.0032 | — | |
| 10 | BUDDY | 0.8756
±
0.0011 | — | |
| 11 | SIGN | 0.8753
±
0.0012 | — | |
| 12 | Neo-GNN | 0.8726
±
0.0084 | — | |
| 13 | GCN | 0.8705
±
0.0004 | — | |
| 14 | GDGNN | 0.8696
±
0.0028 | — | |
| 15 | AML (G) | 0.8670
±
0.0005 | — | |
| 16 | SEAL | 0.8632
±
0.0052 | — | |
| 17 | YinYanGNN | 0.8621
±
0.0009 | — | |
| 18 | AML-R (G) | 0.8591
±
0.0004 | — | |
| 19 | AGDN w/GraphSAINT | 0.8549
±
0.0029 | — | |
| 20 | AGDN | 0.8549
±
0.0029 | — | |
| 21 | LEADING | 0.8520 | — | |
| 22 | PLNLP | 0.8492
±
0.0029 | — | |
| 23 | Best GNN | 0.8490
±
0.0006 | — | |
| 24 | GCN | 0.8490
±
0.0006 | — | |
| 25 | NO | 0.8474
±
0.0021 | — | |
| 26 | GCN | 0.8474
±
0.0021 | — | |
| 27 | Classification | 0.8464 | — | |
| 28 | GPRGNN | 0.8455
±
0.0355 | — | |
| 29 | SuperTMA | 0.8375
±
0.0043 | — | |
| 30 | SAGE+MLP | 0.8367
±
0.0007 | — | |
| 31 | NoFeat | 0.8336
±
0.0014 | — | |
| 32 | RandomTMA | 0.8328
±
0.0024 | — | |
| 33 | Supervised-FT | 0.8310 | — | |
| 34 | ZO-OS | 0.8285
±
0.0043 | — | |
| 35 | SAGE | 0.8260
±
0.0036 | — | |
| 36 | GraphSAGE | 0.8260
±
0.0036 | — | |
| 37 | ZO-S | 0.8245
±
0.0062 | — | |
| 38 | PSGD-PA | 0.8240
±
0.0028 | — | |
| 39 | QGAT | 0.8220
±
0.0127 | — | |
| 40 | GGS | 0.8195
±
0.0020 | — | |
| 41 | LLCG | 0.8188
±
0.0002 | — | |
| 42 | Pre-trained | 0.8180 | — | |
| 43 | GCN+DRNL | 0.8127
±
0.0031 | — | |
| 44 | NeighborSampling | 0.8044
±
0.0010 | — | |
| 45 | GATv2 | 0.8014
±
0.0071 | — | |
| 46 | ClusterGCN | 0.8004
±
0.0025 | — | |
| 47 | GraphSAINT | 0.7985
±
0.0040 | — | |
| 48 | DE-GNN | 0.7885
±
0.0017 | — | |
| 49 | ZO | 0.7850
±
0.0108 | — | |
| 50 | SBERT | 0.7730 | — |
SOTA performance
set a new SOTA at the time
matched the SOTA at the time