OGB
ogbl-wikikg2
Structure
1Graphs
2,500,604Nodes
17,137,181Edges
Coverage
49Results
1Variants
9Papers
- Are Large Language Models Good Temporal Graph Learners? Are Large Language Models Good Temporal Graph Learners?
- TMLR 2024: Expressivity of Representation Learning on Continuous-Time Dynamic Graphs: An Information-Flow Centric Review Expressivity of Representation Learning on Continuous-Time Dynamic Graphs: An Information-Flow Centric Review TMLR 2024
- TripleRE: Knowledge Graph Embeddings via Tripled Relation Vectors TripleRE: Knowledge Graph Embeddings via Tripled Relation Vectors
- NeurIPS 2022: A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs NeurIPS 2022
- ICLR 2022: StarGraph: Knowledge Representation Learning based on Incomplete Two-hop Subgraph StarGraph: Knowledge Representation Learning based on Incomplete Two-hop Subgraph ICLR 2022
- PIE: a Parameter and Inference Efficient Solution for Large Scale Knowledge Graph Embedding Reasoning PIE: a Parameter and Inference Efficient Solution for Large Scale Knowledge Graph Embedding Reasoning
- InterHT: Knowledge Graph Embeddings by Interaction between Head and Tail Entities InterHT: Knowledge Graph Embeddings by Interaction between Head and Tail Entities
- PairRE: Knowledge Graph Embeddings via Paired Relation Vectors PairRE: Knowledge Graph Embeddings via Paired Relation Vectors
- NeurIPS 2020: OGB OGB NeurIPS 2020
Link prediction Predict missing heads or tails of Wikidata triples. Entities are identifiers with learned representations in the standard task; no text or image features are supplied by the core artifact. Evaluated by MRR.
Dataset details
Task
Predict missing heads or tails of Wikidata triples. Filtered MRR ranks each true entity against 500 sampled corrupt heads and 500 sampled corrupt tails that do not form known triples.
Data
The graph contains 2,500,604 retained entities, 535 directed relation types, and 17,137,181 triples extracted from historical Wikidata dumps after filtering rare entities.
Features
Entities are identifiers with learned representations in the standard task; no text or image features are supplied by the core artifact. Relations are typed directed edges.
Splits and evaluation
Triples in the May 2015 snapshot train the model, additions present by August validate it, and additions by November form test. Only entities and relation types already present in May are retained.
Quirks and pitfalls
Wikidata is incomplete and collaboratively edited, so an absent triple is not necessarily false. The predecessor ogbl-wikikg had a negative-sampling bug and is deprecated. Sampled MRR differs from ranking all entities, and inverse-relation augmentation changes the model input.
Sources
SOTA Progression
new SOTA
matches SOTA at that time
other proposed result
other baseline result
Leaderboard
| # | Model | MRR ↑ | Paper |
|---|---|---|---|
| 1 | InterHT + DigPiece | 0.7293
±
0.0018 | |
| 2 | StarGraph + TripleRE' | 0.7290 | |
| 3 | TripleRE + StarGraph | 0.7201
±
0.0011 | |
| 4 | TNCN | 0.7110
±
0.0070 | |
| 5 | InterHT + NodePiece | 0.6779
±
0.0018 | |
| 6 | A*Net | 0.6767 | |
| 7 | TripleRE + NodePiece | 0.6582
±
0.0020 | |
| 8 | Qwen3-8B | 0.6510 | |
| 9 | Qwen2.5-7B | 0.6480 | |
| 10 | ComplEx-RP | 0.6392
±
0.0045 | |
| 11 | ComplEx-RP (50dim) | 0.6392
±
0.0045 | |
| 12 | ComplEx+RP | 0.6392 | |
| 13 | Information-Flow (IF) framework | 0.6110
±
0.0110 | |
| 14 | Qwen3-1.7B | 0.6070 | |
| 15 | TripleREv2 | 0.6045
±
0.0017 | |
| 16 | Mistral-7B-v0.3 | 0.6040 | |
| 17 | Llama3-8B | 0.6040 | |
| 18 | EdgeBank | 0.6000 | |
| 19 | PairRE+LRE | 0.5970 | |
| 20 | PairRE | 0.5805 | |
| 21 | TripleRE | 0.5794
±
0.0020 | |
| 22 | TGN | 0.5770
±
0.0150 | |
| 23 | NodePiece + AutoSF | 0.5703
±
0.0035 | |
| 24 | AutoSF + NodePiece | 0.5703
±
0.0035 | |
| 25 | AutoSF | 0.5458
±
0.0052 | |
| 26 | PairRE | 0.5289
±
0.0030 | |
| 27 | PairRE (200dim) | 0.5208
±
0.0027 | |
| 28 | PairRE | 0.5208 | |
| 29 | ComplEx (250dim) | 0.5027
±
0.0027 | |
| 30 | HTGN (UTG) | 0.4640
±
0.0050 | |
| 31 | TransE | 0.4536
±
0.0030 | |
| 32 | RotatE (250dim) | 0.4332
±
0.0025 | |
| 33 | RotatE | 0.4332
±
0.0025 | |
| 34 | RotatE (5×dim) | 0.4332
±
0.0025 | |
| 35 | TransE (500dim) | 0.4256
±
0.0030 | |
| 36 | TransE (5×dim) | 0.4256
±
0.0030 | |
| 37 | TransE | 0.4256 | |
| 38 | GPT-4.1-mini | 0.4250 | |
| 39 | ComplEx | 0.4028
±
0.0030 | |
| 40 | ComplEx (5×dim) | 0.4027
±
0.0027 | |
| 41 | TGN | 0.3960
±
0.0600 | |
| 42 | GCLSTM (UTG) | 0.3740
±
0.0100 | |
| 43 | DistMult | 0.3729
±
0.0045 | |
| 44 | DistMult (5×dim) | 0.3729
±
0.0045 | |
| 45 | GCN (UTG) | 0.3360
±
0.0090 | |
| 46 | TGAT | 0.2820
±
0.0220 | |
| 47 | DyRep | 0.2630
±
0.0050 | |
| 48 | JODIE | 0.2610
±
0.1230 | |
| 49 | GraphMixer | 0.1180
±
0.0020 |
SOTA performance
set a new SOTA at the time
matched the SOTA at the time