Open Graph Benchmark - MAG (Academic Graph)

Large heterogeneous graph of academic citations and collaborations from Microsoft Academic Graph. Contains 4 node types: papers, authors, institutions, and fields of study. 5 edge types: author-paper (writes), author-institution (affiliated-with), paper-venue (published-in), and topic edges. Used extensively for heterogeneous graph neural network benchmarking and node classification tasks.

npx keepset fetch ogbn-mag

Nodes: 1,939,743
Edges: 21,111,007
Size: 411.8 MiB

Fetch and preview this dataset on this page, or download the result as JSON. The command line fetches it without a browser.

Overview

Id	`ogbn-mag`
Category	`graph/v1`
Kind	graph
Domain	Academic heterogeneous
Licence	ODC-By-1.0
Catalogue	`keepset` 1.33.0

Sources and provenance

#	URL	Status	Retrieved	Read by recipe	Pinned snapshot	SHA-256
0	https://data.pyg.org/datasets/ogbn_mag.zip	live	2026-03-26
1	https://snap.stanford.edu/ogb/data/nodeproppred/mag.zip	live	2026-10-21	yes	20261022042203raw	`2afe62ead87f`

Citations

Hu et al. (2020). Open Graph Benchmark: Datasets for Machine Learning on Graphs. NeurIPS.

Recipe

Source	#1
Archive	zip
Files	matching `[^/]+/raw/(?:num-node-list\|num-node-dict\|edge\|num-edge-list\|triplet-type-list\|relations/[^/]+/(?:edge\|num-edge-list))\.csv\.gz`
Inner archive	gzip
Parser format	`"ogb"`
Parser encoding	`"utf-8"`

Crosswalk

keepset/topology to graph/v1, as a crossmap crosswalk.

Rule	From	To	Direction
move	`meta`	`meta`	both
move	`nodes[].id`	`nodes[].id`	both
move	`edges[].source`	`edges[].source`	both
move	`edges[].target`	`edges[].target`	both

Expected counts

Edges	21,111,007
Nodes	1,939,743

Properties

Components	1
Connected	yes
Cyclic	yes
Density	0.00001121
Directed	yes
Download bytes	411.8 MiB
Edge types	1
Heterogeneous	yes
Multigraph	no
Node types	2
Raw bytes	0 B
Self loops	no
Weighted	no