The structure of scientific collaboration networks

• Milgram’s experiment found about six steps between randomly selected people and helped establish the small-world hypothesis.The experiment nevertheless provided limited information about detailed social-network structure.Source: §I · p. 1
• Existing friendship-network studies are labor-intensive, typically small, and sensitive to subjective judgments about who counts as an acquaintance.These limitations restrict their usefulness for understanding detailed network structure.Source: §I · p. 1
• Power grids, the internet, and air-traffic networks provide abundant data but do not directly measure contact between people.They are therefore poor proxies for actual human acquaintance patterns.Source: §I · pp. 1–2
• Movie-actor networks contain many people, but appearing in the same film does not reliably establish meaningful acquaintance or everyday interaction.Using movie coappearance to infer ordinary human relationships can therefore be misleading.Source: §I · p. 2
• Scientific collaboration offers a large human network with a precise acquaintance definition based on coauthored papers.The study describes a network containing over a million people.Source: §I · p. 2

• Two scientists are connected when they have coauthored a paper, providing a stringent but consistently applicable definition of scientific acquaintance.The definition excludes scientists who know one another but have never collaborated on a paper.Source: §II · p. 2
• Collaboration graphs cover biomedical research, theoretical and high-energy physics, and computer science using four databases and papers or preprints from 1995–1999.Database sizes range from 13,000 papers in NCSTRL to 2 million in MEDLINE.Source: §II · p. 2
• The publication record supports bibliographical proximity measures such as the Erdős number, which counts coauthorship steps from Paul Erdős.Authors who published with Erdős have number 1, while those connected through one coauthor have number 2.Source: §II · p. 2
• The study measures collaborators, papers, and clustering, defined as the probability that two collaborators of one scientist have collaborated with each other.Source: §II · p. 2

III. RESULTS

• The analysis summarizes seven networks, including four complete databases and three subject-specific Los Alamos subsets, using results compiled in Table I.The databases cover biomedical research, physics, high-energy physics, and computer science.Source: §III · p. 2, Table I · p. 3

a. Number of authors

• Author-count estimates use all-initial and first-initial name matching to provide upper and lower bounds for database size and related quantities.The two analyses respectively risk splitting one author into several identities or conflating distinct authors.Source: §III · p. 2

b. Mean papers per author and authors per paper

• Average collaboration size differs sharply by field: theoretical databases average about four collaborators, astrophysics reaches 15, and experimental high-energy physics reaches 173 per author.The SPIRES database includes a paper with 1681 authors, reflecting the scale of experimental collaborations.Source: §III · p. 3

c. Number of collaborators

• Collaborator and paper-count distributions are well fit by power laws with exponential cutoffs, with fits achieving R2 > 0.99 and p-values below 10^-3.The cutoff may reflect the five-year observation window or scientists’ finite working lifetimes, and cutoff sizes vary across fields.Source: §III · p. 4
• The collaborator-distribution exponent τ is near 1 in hard sciences but 2.5 in MEDLINE, so highly connected scientists dominate averages in physics and computer science but not biomedicine.The paper identifies τ = 2 as the dividing line between these two regimes.Source: §III · p. 4

d. The giant component

• The largest connected component contains roughly 80–90% of authors in most databases, exceeding half the network even in high-energy theory and computer science.The much smaller second-largest components indicate that these networks are well inside the percolating regime rather than near fragmentation.Source: §III · p. 4

e. Average degrees of separation

• Typical scientist-to-scientist distance is about six steps, while measured distances scale logarithmically with community size relative to random-graph expectations.The reported correlation with log N/log z is R2 = 0.83; network diameters are generally about 20 among connected pairs.Source: §III · p. 5

f. Clustering

• Clustering is strong: two scientists typically have at least a 30% probability of collaborating when they have both collaborated with a third scientist.The effect cannot be explained entirely by multi-author papers; collaborator introductions or institutions may also generate new collaborations, while MEDLINE has substantially lower clustering than hard-science databases.Source: §III · pp. 5–6

• Average separation is typically five or six collaborator steps and scales logarithmically with community size across the studied scientific fields.Source: §IV · p. 6
• Scientific collaboration networks are highly clustered, so scientists sharing a collaborator are much more likely to collaborate with each other than randomly chosen scientists.The authors suggest that scientists introducing collaborators to one another may contribute to this pattern.Source: §IV · p. 6
• Collaborator and paper-count distributions are well fit by power laws with exponential cutoffs, possibly because the study covers only a finite time window.Source: §IV · p. 6
• Collaboration patterns differ substantially across fields: experimental high-energy physics has far more collaborators per author, while biomedicine has lower clustering and is the only field with a collaborator-distribution exponent above 2.In biomedicine, average network properties are dominated by many scientists with few collaborators rather than a few with many.Source: §IV · p. 6
• The study is explicitly a first look, leaving many additional theoretical measures of socially important network structure for future analysis.Source: §IV · p. 6