Learning and the structure of citation networks

The distribution of citations received by scientific publications can be approximated by a power law, a finding that has been explained by “cumulative advantage”. This paper argues that socially embedded learning is a plausible mechanism behind this cumulative advantage. A model assuming that scientists face a time trade-off between learning and writing papers, that they learn the papers known by their peers, and that they cite papers they know, generates a power law distribution of popularity, and a shifted power law for the distribution of citations received. The two distributions flatten if there is relatively more learning. The predicted exponent for the distribution of citations is independent of the average in-(or out-) degree, contrary to an untested prediction of the reference model (Price, 1976). Using publicly available citation networks, an estimate of the share of time devoted to learning (against producing) is given around two thirds.

Keywords: shifted power law, scale free networks, two-mode networks, cumulative advantage, polynomial attachment kernel, innovation, diffusion.

JEL codes: D83, D85, O31, O33