README.md

metadata

pretty_name: PrimeKG-CL Benchmark v1.0 (Anon)
license: cc-by-4.0
task_categories:
  - other
language:
  - en
tags:
  - biomedical
  - knowledge-graph
  - continual-learning
  - link-prediction
size_categories:
  - 10M<n<100M

PrimeKG-CL Benchmark v1.0 (Anon)

PrimeKG-CL is a continual graph learning benchmark built from an evolving biomedical knowledge graph with real temporal change.

Dataset Summary

• Two temporal snapshots (t0, t1) of a biomedical KG.
• Temporal diff data for added/removed/persistent facts.
• 10 continual-learning tasks with train/valid/test splits.
• Stratified evaluation metadata for retention and forgetting analysis.
• Multimodal node features (text + molecular + graph tensors).

Snapshots

• snapshots/kg_t0.csv — June 2021 PrimeKG release: 8,100,498 triples, 129,375 nodes, 30 relations.
• snapshots/kg_t1.csv — July 2023 reconstruction from nine upstream databases (Bgee, CTD, GO, Gene2GO, HPO, HPOA, MONDO, Uberon, HGNC): 13,001,666 triples, 134,211 nodes, 25 relations.
• snapshots/t1_build_info.json — provenance and database versions for the t1 reconstruction.

Temporal diff (t0 -> t1)

• Added: 5,760,234 triples
• Removed: 888,848 triples
• Persistent: 7,208,624 triples

Detailed breakdown is in diffs/diff_t0_t1.json.

Data Structure

Continual tasks

Tasks are in tasks/task_*. Each task directory contains:

• train.txt
• valid.txt
• test.txt

Each line is a tab-separated triple:

• head_id<TAB>relation<TAB>tail_id

Split ratio: 70/10/20.

Task sizes:

• task_0_base: 8,100,498 triples (full t0)
• task_1_disease_related: 17,009
• task_1_drug_related: 125,343
• task_2_disease_related: 115,382
• task_2_gene_protein: 2,850,593
• task_3_gene_protein: 99,761
• task_3_phenotype_related: 47,997
• task_4_biological_process: 116,118
• task_4_phenotype_related: 57,390
• task_5_anatomy_pathway: 2,752,675

Stratified evaluation

test_stratification.json partitions each task's test triples into:

• persistent (in both snapshots)
• added (new in t1)
• removed (only in t0)

This supports stratified-MRR evaluation of retention vs forgetting.

Multimodal features

• features/text_embeddings.pt — BiomedBERT [CLS] embeddings projected to 256-d (36% entity coverage).
• features/mol_features.pt — Morgan fingerprints (radius 2, 1024 bits) for drugs with SMILES (4% entity coverage).
• features/edge_index.pt, features/edge_type.pt — R-GCN tensors for the full t0 graph.
• features/node_has_text.pt, features/node_has_mol.pt — feature availability masks.
• features/node_index_map.csv — global entity_id <-> row_index mapping.
• features/vocab_sizes.json — entity/relation vocabulary sizes.

Files

• LICENSE — MIT (code) + CC BY 4.0 (data)
• croissant.json — MLCommons Croissant metadata
• README.md — dataset card
• statistics.json — benchmark summary stats
• test_stratification.json — per-task persistent/added/removed counts
• diffs/diff_t0_t1.json — full temporal diff
• snapshots/ — kg_t0.csv, kg_t1.csv, provenance metadata
• tasks/ — continual-learning task splits
• features/ — multimodal node features

Intended Uses

• Continual knowledge graph completion
• Continual link prediction
• Retention/forgetting analysis under temporal KG evolution
• Multimodal CGL ablations using structure/text/molecular signals

Limitations

• Dataset quality depends on upstream biomedical resource quality and curation biases.
• Restrictively licensed sources (DrugBank, UMLS, DrugCentral, SIDER, DisGeNET) were not re-queried for t1; corresponding t0 edges were carried forward unchanged.
• This benchmark is for machine learning research and is not intended for clinical decision-making.

Licensing

• Data files: CC BY 4.0 (inheriting PrimeKG terms)
• Code/configuration artifacts: MIT

Users are responsible for compliance with applicable upstream database licenses.

Citation

@inproceedings{primekgcl2026,
  title={PrimeKG-CL: A Continual Graph Learning Benchmark on Evolving Biomedical Knowledge Graphs},
  author={Anonymous},
  booktitle={NeurIPS Datasets and Benchmarks Track},
  year={2026}
}