Patent Classification on USPTO-70k
53.1Micro F1PHAGE
Evaluation results
| Method | Links | ||
|---|---|---|---|
| PHAGEModel category=Claim-level graph, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 53.1 | 34.8 | |
| PatEmbed-BaseModel category=Citation-aware, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 52.3 | 31.7 | |
| PaECTERModel category=Citation-aware, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 51.1 | 33.3 | |
| BERT-for-PatentsModel category=Sequence-only, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 49.1 | 28.8 | |
| PatentSBERTaModel category=Citation-aware, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 45.6 | 25.3 | |
| PatEmbed-LargeModel category=Citation-aware, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 41.5 | 21 | |
| BERT-baseModel category=Sequence-only, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 38.6 | 16.3 | |
| SciBERTModel category=Sequence-only, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg.2026.05 | 34.7 | 13.3 | |
| GraphSAGE + BERT-for-PatentsModel category=Graph-based, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg., Backbone=BERT-for-Patents [CLS], Inference setting=validation-tuned k-NN2026.05 | 33.2 | 23 | |
| GAT + BERT-for-PatentsModel category=Graph-based, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg., Backbone=BERT-for-Patents [CLS], Inference setting=validation-tuned k-NN2026.05 | 32.7 | 21.3 | |
| GCN + BERT-for-PatentsModel category=Graph-based, Evaluation protocol=linear probing, Encoder configuration=frozen encoder, Statistical aggregation=5-run avg., Backbone=BERT-for-Patents [CLS], Inference setting=validation-tuned k-NN2026.05 | 31.2 | 20.5 |