Overview of OGB-LSC

Why a Large-Scale Graph ML Datasets?

Machine Learning (ML) on graphs has attracted immense attention in recent years because of the prevalence of graph-structured data in real-world applications. Modern application domains include web-scale social networks, recommender systems, hyperlinked web documents, knowledge graphs (KGs), as well as molecule simulation data generated by the ever-increasing scientific computation. These domains involve large-scale graphs with billions of edges or a dataset with millions of graphs. Deploying accurate graph ML at scale will have a huge practical impact, enabling better recommendation results, improved web document search, more comprehensive KGs, and accurate ML-based drug and material discovery. However, community efforts to advance state-of-the-art in large-scale graph ML have been extremely limited. In fact, most of graph ML models have been developed and evaluated on extremely small datasets.

Handling large-scale graphs is challenging, especially for state-of-the-art expressive Graph Neural Networks (GNNs) because they make prediction on each node based on the information from many other nodes. Effectively training these models at scale requires sophisticated algorithms that are well beyond standard SGD over i.i.d. data. More recently, researchers improve model scalability by significantly simplifying GNNs, which inevitably limits their expressive power.

However, in deep learning, it has been demonstrated over and over again that one needs big expressive models and train them on big data to achieve the best performance. In graph ML, the trend has been the opposite—models get simplified and less expressive to be able to scale to large graphs. Thus, there is a massive opportunity to move the community to work with realistic and large-scale graph datasets and move the state of the field forward to where it needs to be.

Overview of OGB-LSC

There are three OGB-LSC datasets: MAG240M, WikiKG90Mv2, and PCQM4Mv2, that are unprecedentedly large in scale and cover prediction at the level of nodes, links, and graphs, respectively. An illustrative overview of the three OGB-LSC datasets is provided below.

  • MAG240M is a heterogeneous academic graph, and the task is to predict the subject areas of papers situated in the heterogeneous graph (node classification).
  • WikiKG90Mv2 is a knowledge graph, and the task is to impute missing triplets (link prediction).
  • PCQM4Mv2 is a quantum chemistry dataset, and the task is to predict an important molecular property, the HOMO-LUMO gap, of a given molecule (graph regression).

For each dataset, we carefully design its prediction task and data split so that achieving high prediction performance on the task will have direct impact on the corresponding application. Further details are provided in each dataset page. The dataset statistics as well as basic information are summarized below, showing that our datasets are extremely large.

Task category Name Package #Graphs #Total nodes #Total edges Task Type Metric Download size
Node-level MAG240M >=1.3.2 1 244,160,499 1,728,364,232 Multi-class classification Accuracy 167GB
Link-level WikiKG90Mv2 >=1.3.3 1 91,230,610 601,062,811 KG completion MRR 89GB
Graph-level PCQM4Mv2 >=1.3.2 3,746,619 52,970,652 54,546,813 Regression MAE 59MB

: The PCQM4Mv2 dataset is provided in the SMILES strings. After processing them into graph objects, the eventual file size will be around 8GB.

All of these datasets can be downloaded and prepared using our ogb Python package. The model evaluation and test submission file preparation are also handled by our package. The usage is described in each dataset page. Please install/update it by:

pip install -U ogb
# Make sure below prints the required package version for the dataset you are working on.
python -c "import ogb; print(ogb.__version__)"

In our paper, we further perform an extensive baseline analysis on each dataset, implementing simple baseline models as well as advanced expressive models at scale. We find that advanced expressive models, despite requiring more efforts to scale up, do benefit from large data and significantly outperform simple baseline models that are easy to scale. All of our baseline code is made publicly available to facilitate public research.

Finally, we provide the leaderboard infrastructure (see here for more details) to evaluate models’ prediction on the hidden test sets. There are two kinds of test sets: (1) test-dev and (2) test-challenge. (1) test-dev is used for the public leaderboards, where the community can evaluate their model performance throughout the year (with some restriction on the frequency of submissions). (2) test-challenge will be held out for future ML competitions and will be touched only during the competition.

Next Steps

- Learn more about the datasets
- View/Contribute to leaderboards


Details about our datasets and our initial baseline analysis are described in our OGB-LSC paper. If you use OGB-LSC in your work, please cite our paper (Bibtex below)

  title={OGB-LSC: A Large-Scale Challenge for Machine Learning on Graphs},
  author={Hu, Weihua and Fey, Matthias and Ren, Hongyu and Nakata, Maho and Dong, Yuxiao and Leskovec, Jure},
  journal={arXiv preprint arXiv:2103.09430},