yradwan147/PrimeKGCL

PrimeKG-CL

PrimeKG-CL is a continual graph learning benchmark built on an evolving biomedical knowledge graph. It contains two temporal snapshots (t0, t1), temporal diffs, 10 continual-learning tasks, and multimodal node features for evaluating retention/forgetting under distribution shift.

This dataset repository stores the benchmark under the benchmark/ directory.

Dataset Summary

• Domain: Biomedical knowledge graphs
• Primary use: Continual knowledge graph completion / link prediction
• Snapshots:
  • benchmark/snapshots/kg_t0.csv (June 2021 PrimeKG release): 8,100,498 triples, 129,375 nodes, 30 relations
  • benchmark/snapshots/kg_t1.csv (July 2023 reconstruction): 13,001,666 triples, 134,211 nodes, 25 relations
• Temporal change (t0 -> t1):
  • Added: 5,760,234 triples
  • Removed: 888,848 triples
  • Persistent: 7,208,624 triples

Repository Structure

• benchmark/README.md: benchmark documentation
• benchmark/LICENSE: benchmark licensing terms
• benchmark/snapshots/: temporal KG snapshots and build provenance
• benchmark/diffs/diff_t0_t1.json: full temporal diff and relation-level breakdown
• benchmark/tasks/task_*/: 10 continual tasks with train.txt, valid.txt, test.txt
• benchmark/test_stratification.json: persistent/added/removed test strata per task
• benchmark/features/: multimodal features and graph tensors
• benchmark/statistics.json: benchmark-level summary statistics
• benchmark/croissant.json: MLCommons Croissant metadata

Data Format

• Triple files are tab-separated text with one triple per line:
  • head_id<TAB>relation<TAB>tail_id
• Task splits use 70/10/20 train/valid/test.
• Feature tensors are stored as .pt files.

Suggested Usage

1. Use task_0_base as the base pretraining stage on t0.
2. Train/evaluate sequentially on tasks task_1_* through task_5_*.
3. Report both aggregate and stratified metrics (persistent vs added vs removed), using benchmark/test_stratification.json.

Limitations and Considerations

• This benchmark reflects biomedical database curation quality and biases from source resources.
• t1 reconstruction intentionally excludes re-querying some restrictively licensed sources; corresponding t0 edges are carried forward unchanged.
• Performance on this benchmark does not imply clinical validity.

Licensing

• Data files: CC BY 4.0 (inherits PrimeKG license)
• Code/config components: MIT

Users must comply with applicable upstream database licenses where required.

Citation

If you use this dataset, please cite:

@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}
}