anonym124/primekgcl-benchmark-v1-anon

PrimeKG-CL是一个持续图学习基准，基于具有真实时间变化的生物医学知识图谱构建。它包括两个时间快照（t0和t1）的生物医学知识图谱，时间差异数据用于添加/删除/持久事实，10个具有训练/验证/测试分割的持续学习任务，用于保留和遗忘分析的分层评估元数据，以及多模态节点特征（文本+分子+图张量）。快照包括2021年6月的PrimeKG发布（8,100,498个三元组，129,375个节点，30个关系）和2023年7月从九个上游数据库重建的版本（13,001,666个三元组，134,211个节点，25个关系）。时间差异显示添加了5,760,234个三元组，删除了888,848个三元组，持久了7,208,624个三元组。任务目录包含每个任务的训练、验证和测试文件，每行是制表符分隔的三元组。分层评估将测试三元组分为持久、添加和删除三类，以支持保留与遗忘的分层MRR评估。多模态特征包括文本嵌入、分子指纹和R-GCN张量等。数据集旨在用于持续知识图谱补全、持续链接预测、时间知识图谱演化下的保留/遗忘分析，以及使用结构/文本/分子信号的多模态CGL消融研究。PrimeKG-CL is a continual graph learning benchmark built from an evolving biomedical knowledge graph with real temporal change. It includes 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, and multimodal node features (text + molecular + graph tensors). Snapshots cover June 2021 PrimeKG release (8,100,498 triples, 129,375 nodes, 30 relations) and July 2023 reconstruction from nine upstream databases (13,001,666 triples, 134,211 nodes, 25 relations). Temporal diff shows 5,760,234 added triples, 888,848 removed triples, and 7,208,624 persistent triples. Task directories contain train/valid/test files per task, each line as a tab-separated triple. Stratified evaluation partitions test triples into persistent, added, and removed categories for stratified-MRR evaluation of retention vs forgetting. Multimodal features include text embeddings, molecular fingerprints, and R-GCN tensors. The dataset is intended for continual knowledge graph completion, continual link prediction, retention/forgetting analysis under temporal KG evolution, and multimodal CGL ablations using structure/text/molecular signals.