Link Prediction on WikiKG90M v2 (val)
0.301MRRKGT5-context, T5 base
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| KGT5-context, T5 baseParams=220M2023.05 | 0.301 | — | |
| DGLKE + Rule MiningParams=18.2B2023.05 | 0.292 | — | |
| BESSParams=23.3B2023.05 | 0.292 | — | |
| PIE-RMParams=18.2B2023.05 | 0.254 | — | |
| KGT5Params=60M2022.03 | 0.221 | — | |
| KGT5, T5 smallParams=60M2023.05 | 0.221 | — | |
| TransE-ConcatParams=18.2B2022.03 | 0.206 | — | |
| TransE-ConcatParams=18.2B2023.05 | 0.206 | — | |
| ComplEx-ConcatParams=18.2B2022.03 | 0.205 | — | |
| ComplEx-ConcatParams=18.2B2023.05 | 0.205 | — | |
| ComplExParams=18.2B2023.05 | 0.182 | — | |
| ComplEx-MPNetParams=307K2022.03 | 0.126 | — | |
| ComplExParams=18.2B2022.03 | 0.115 | — | |
| TransE-MPNetParams=307K2022.03 | 0.113 | — | |
| TransEParams=18.2B2022.03 | 0.11 | — | |
| TransEParams=18.2B2023.05 | 0.11 | — | |
| ComplEx2026.03 | — | 0.1649 | |
| Direct ensemblestrategy=grid search above all models with non-zero weights2026.03 | — | 0.28 | |
| Ensemble with Model Selectionstrategy=first select best models (TransE-0,TransE-1,TransE-2,ComplEx, OTE-0, OTE-2) then grid search on model weights2026.03 | — | 0.2839 | |
| NOTE-02026.03 | — | 0.1561 | |
| NOTE-1negative sample size=12002026.03 | — | 0.1648 | |
| NOTE-2embeddings=neighbor enhanced embeddings and randomly initialized embeddings w/ MLP layer2026.03 | — | 0.1654 | |
| NOTE-3negative sample size=1200, embeddings=neighbor enhanced embeddings and randomly initialized embeddings w/ MLP layer2026.03 | — | 0.1592 | |
| TransE-02026.03 | — | 0.214 | |
| TransE-1batch size=20480, negative sample size=204802026.03 | — | 0.2094 | |
| TransE-2batch size=20480, negative sample size=20480, embeddings=MPNet and randomly initialized embeddings w/ MLP layer2026.03 | — | 0.2114 | |
| TransE-3embeddings=neighbor enhanced embeddings and randomly initialized embeddings w/ MLP layer2026.03 | — | 0.1877 |