ogbl-collab, ogbl-ppa, ogbl-citation2
收藏资源简介:
本研究涉及的数据集主要用于评估图神经网络在链接预测任务中的性能,包括ogbl-collab、ogbl-ppa和ogbl-citation2。这些数据集用于评估新模型的有效性,并揭示了当前评估方法中存在的多个陷阱,如基准性能低于实际、缺乏统一的数据分割和评估指标、使用不切实际的评估设置等。研究通过实施公平比较和创建更实际的评估设置,推动了链接预测领域的挑战和机遇,使其更符合现实世界的情况。
The datasets involved in this study are primarily used to evaluate the performance of graph neural networks (GNNs) on link prediction tasks, including ogbl-collab, ogbl-ppa, and ogbl-citation2. These datasets serve to evaluate the efficacy of novel models, and they uncover several pitfalls in current link prediction evaluation workflows, including baseline performance lower than real-world actual levels, lack of unified data splits and evaluation metrics, and employment of unrealistic evaluation settings, among others. By conducting fair comparative studies and establishing more realistic evaluation setups, this study advances the challenges and opportunities in the link prediction field, making the domain more consistent with real-world practical scenarios.

- 1Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New Benchmarking密歇根州立大学 · 2023年



