ogbl-collab

Rank future author-author collaboration edges above sampled non-edges, given collaborations observed in earlier years. The official score is Hits@50 against 100,000 sampled negative pairs per positive edge.

The undirected graph is a MAG subset with 235,868 author nodes. Multiple collaborations between the same authors in different years remain separate temporal edges; edge weight records the number of jointly authored papers in that year.

Each author has a 128-dimensional vector obtained by averaging word embeddings over that author’s papers. Edges carry year and collaboration-count metadata.

Edges through 2017 are training data, 2018 collaborations validate the model, and 2019 collaborations form test. OGB permits adding validation edges when making test predictions, so evaluation code must state which historical graph is used.

Sampled-negative Hits@50 is not equivalent to ranking every absent edge, and absence of an edge does not prove two authors will never collaborate. Treating yearly parallel edges as one static edge loses timing and weight information. Random edge splits leak future collaboration patterns.