A Memory-Theoretic Account of Citation Propagation

Abstract

Despite the common assumption that citations are indicative of an article’s scientific merit, increasing evidence indicates that citation counts are largely driven by variables unrelated to quality. In this paper, we demonstrate that, in fact, distributions of citation counts are consistent with a stochastic process based on the exposure probability of a paper. The proposed exposure model anticipates that small alterations in the exposure of scientific papers early on in their life cycle are magnified over time and can lead to the emergence of highly cited papers. This effect occurs even when there is no variation in the starting point exposure probabilities (i.e., when assuming a level playing field where all articles are treated equally and of equal ‘quality’), and is exacerbated when a small number of articles have greater exposure early on. The model anticipates the emergence of ‘black swan’ type articles - articles that garner large quantities of citations through random processes.

A memory-theoretic account of citation propagation

“…The entire process is repeated for c epochs to simulate the evolution of citation patterns across the life cycle of all 100 articles. Individuals interested in exploring the predictions of the model can access R code to run the model at https://osf.io/hmbup/ [45].…”

“…The models were fit to R = 46 690 articles (cited C = 2 161 838 times) from 32 consecutive years (1985–2016) of articles included in the previously described database from the psychological and brain sciences [49,50]. The citation data for each of these articles was obtained using the scholar [51] package in R [52] and is available at [45], along with analyses for all 32 years of data. Citation data were gathered in January 2019 using the scholar API tool; citation counts for 3929 articles were checked manually by the first author to verify the accuracy of the automated tool.…”

“…A more detailed description of the model is included in appendix A. This supplemental material along with Python code for running the model is provided on [ 45 ]. Computer code used to distill the model output and produce the plots included in this article was written in R and is also provided at [ 45 ].…”