Algorithm and System Co-design for Efficient Subgraph-based Graph Representation Learning

Following the format of the Open Graph Benchmark (OGB), we design four prediction tasks of relations (mag-write, mag-cite) and higher-order patterns (tags-math, DBLP-coauthor) and construct the corresponding datasets over heterogeneous graphs and hypergraphs [1]. The original ogb-mag dataset only contains features for 'paper'-type nodes. We add the node embedding provided by [2] as raw features for other node types in MAG(P-A)/(P-P). For these four tasks, the model is evaluated by one positive query paired with a certain number of randomly sampled negative queries (1:1000 by default, except for tags-math 1:100).