MAG for Heterogeneous Graph Learning

We provide an academic graph based on a snapshot of the Microsoft Academic Graph from 26.05.2021. The Microsoft Academic Graph (MAG) is a large-scale dataset containing information about scientific publication records, their citation relations, as well as authors, affiliations, journals, conferences and fields of study. We acknowledge the Microsoft Academic Graph using the URI https://aka.ms/msracad. For more information regarding schema and the entities present in the original dataset please refer to: MAG schema. 

MAG for Heterogeneous Graph Learning
We use a recent version of MAG from May 2021 and extract all relevant entities to build a graph that can be directly used for heterogeneous graph learning (node classification, link prediction, etc.). The graph contains all English papers, published after 1900, that have been cited at least 5 times per year since the time of publishing. For fairness, we set a constant citation bound of 100 for papers published before 2000. We further include two smaller subgraphs, one containing computer science papers and one containing medicine papers. 

Nodes and features
We define the following nodes:

• paper with mag_id, graph_id, normalized title, year of publication, citations and a 128-dimension title embedding built using word2vec
  No. of papers: 5,091,690 (all), 1,014,769 (medicine), 367,576 (computer science);
• author with mag_id, graph_id, normalized name, citations
  No. of authors: 6,363,201 (all), 1,797,980 (medicine), 557,078 (computer science);
• field with mag_id, graph_id, level, citations denoting the hierarchical level of the field where 0 is the highest-level (e.g. computer science)
  No. of fields: 199,457 (all), 83,970 (medicine), 45,454 (computer science);
• affiliation with mag_id, graph_id, citations
  No. of affiliations: 19,421 (all), 12,103 (medicine), 10,139 (computer science); 
• venue with mag_id, graph_id, citations, type denoting whether conference or journal
  No. of venues: 24,608 (all), 8,514 (medicine), 9,893 (computer science). 

Edges
We define the following edges:

• author is_affiliated_with affiliation
  No. of author-affiliation edges: 8,292,253 (all), 2,265,728 (medicine), 665,931 (computer science);
• author is_first/last/other paper
  No. of author-paper edges: 24,907,473 (all), 5,081,752 (medicine), 1,269,485 (computer science);
• paper has_citation_to paper
  No. of author-affiliation edges: 142,684,074 (all), 16,808,837 (medicine), 4,152,804 (computer science);
• paper conference/journal_published_at venue
  No. of author-affiliation edges: 5,091,690 (all), 1,014,769 (medicine), 367,576 (computer science);
• paper has_field_L0/L1/L2/L3/L4 field
  No. of author-affiliation edges: 47,531,366 (all), 9,403,708 (medicine), 3,341,395 (computer science);
• field is_in field
  No. of author-affiliation edges: 339,036 (all), 138,304 (medicine), 83,245 (computer science);

We further include a reverse edge for each edge type defined above that is denoted with the prefix rev_ and can be removed based on the downstream task. 

Data structure
The nodes and their respective features are provided as separate .tsv files where each feature represents a column. The edges are provided as a pickled python dictionary with schema:

{target_type:
    {source_type:
        {edge_type:
            {target_id:
                {source_id:
                    {time
                    }
                }
            }
        }
    }
}

We provide three compressed ZIP archives, one for each subgraph (all, medicine, computer science), however we split the file for the complete graph into 500mb chunks. Each archive contains the separate node features and edge dictionary.