Scientometrics Computer Science Applications
Network position and scientific recognition: author centrality and citation outcomes in top AI conferences
Abstract
This study examines whether authors’ positions in collaboration networks are associated with citation outcomes in top artificial-intelligence conferences. To capture complementary dimensions of collaboration-network position, we propose three centrality metrics: HCTCD, TDDH, and ETA. The empirical framework uses year-specific citation percentiles, the Smithson–Verkuilen boundary adjustment, Probit-link beta regression, and 8-year pre-publication collaboration networks. After title–year–venue deduplication and removal of the 2015–2016 calibration years, 3,573 papers form the primary evaluation sample. Parameters for each proposed centrality metric are selected on earlier calibration or development cohorts and fixed before the corresponding evaluations. In the primary evaluation sample, HCTCD is positively associated with citation standing and yields fit comparable to conventional closeness. Author-order-weighted average aggregation has the lowest observed AIC among seven HCTCD aggregation specifications, with virtually identical fit to the unweighted average. TDDH and ETA provide later-cohort extensions, and both reduce mean squared prediction error relative to controls-only models across four prediction models. Double machine learning provides an orthogonalization sensitivity analysis of the association between author centrality and citation outcomes. Papers without usable author identifiers differ observably from included papers, while inverse-probability-weighted estimates adjusting for these observed differences remain close to the unweighted estimates. Overall, pre-publication network position adds information about scientific recognition, while temporal, path, and team-reach adjustments differ in value across empirical settings.