Source-linked AI summary
Predicting Scientific Success Based on Coauthorship Networks
Emre Sarigöl, Rene Pfitzner, Ingo Scholtes, Antonios Garas, Frank Schweitzer
TL;DR
The paper asks to what extent scientific article success reflects social influence rather than only the work itself. It analyzes coauthorship-network centrality and uses machine learning to predict five-year citation success. The results show strong statistical dependence between author centrality and citation success, while cautioning that the relationship is not necessarily causal.
Problem
The paper examines whether citation-based scientific success, often treated as an objective measure, is related to authors’ social structures and coauthorship-network positions.
Method
Using more than 100,000 publications, the study compares centrality distributions and trains a Random Forest classifier on coauthorship-network centrality measured at publication.
Results
60% precision was achieved when predicting whether an article belonged to the 10% most cited articles, and highly cited papers had authors with significantly higher average centrality.
Takeaways & Limitations
Citation measures should be interpreted as reflecting multiple influencing factors, including social influence, rather than as socially unbiased measures of scientific success.
Takeaways & Limitations
The observed association between centrality and citation numbers does not establish direct causality, because both could reflect scientific excellence or another underlying factor.
Abstract
from arXiv · showhide
We address the question to what extent the success of scientific articles is due to social influence. Analyzing a data set of over 100000 publications from the field of Computer Science, we study how centrality in the coauthorship network differs between authors who have highly cited papers and those who do not. We further show that a machine learning classifier, based only on coauthorship network centrality measures at time of publication, is able to predict with high precision whether an article will be highly cited five years after publication. By this we provide quantitative insight into the social dimension of scientific publishing - challenging the perception of citations as an objective, socially unbiased measure of scientific success.
1 Introduction
Citation metrics are widely used as objective proxies for scientific success, yet scientific activity may remain socially influenced. The paper studies whether coauthorship-network position predicts citation success.
- Citation counts, citations per article, and the h-index are used to quantify research impact and assess scientific success.
- Scientific activities may be influenced by social aspects, even when citation-based evaluation is intended to be free from personal beliefs, biases, or social influence.
- The study analyzes more than 100,000 publications by more than 160,000 authors using time-evolving coauthorship networks as a proxy for social structure.
- A Random Forest classifier using complementary network-centrality measures predicts membership among the 10% most cited articles with 60% precision.
2 The Complex Character of Citations
Citations have complex social and disciplinary dimensions beyond simple attribution of credit. Prior work links citation behavior to collaboration structures, while this paper examines authors’ network centrality and citation success.
- A complete theory of citations is still missing, and citations can express discursive relations between articles and professional relations between authors.
- Citation practices differ significantly across scientific disciplines, complicating the definition of universal citation-based impact measures.
- Article citation counts correlate with the number of authors and institutions involved in producing the article.
- Prior network-based studies incorporated collaboration-network distance into citation indexes and extended self-citation concepts using topological distance.
- This paper differs by studying correlations between authors’ collaboration-network centrality and the citation success of their research articles.
3 Time-Evolving Collaboration and Citation Networks
The study constructs time-evolving computer-science coauthorship networks from Microsoft Academic Search data and evaluates whether authors’ network positions predict later citation success. It uses network centrality measured around publication and citations accumulated over five years.
- The dataset contains more than 100,000 computer-science articles published between 1996 and 2008, sampled through 1,000 authors selected from 20,000 highly ranked authors.
- Publication records identify 160,891 researchers, while coauthorship-network nodes represent authors and links represent coauthor relationships.
- Two-year coauthorship windows slide in one-year increments, producing 11 time slices representing collaboration structures between 1996 and 2008.
- The analysis uses the largest connected component and computes degree, eigenvector, betweenness, and k-core centrality measures for authors.
- For each publication, the hindcasting approach measures author centrality in the preceding two-year network and counts citations accumulated during the five years after publication.
- A paper is classified as successful when its five-year citation count exceeds that of 90% of papers published in the same year.
4 Statistical Dependence of Coauthorship Structures and Citations
The paper tests whether authors’ coauthorship-network centrality is associated with citation success and whether citation-success changes relate to later centrality. Using centrality comparisons, conditional probabilities, and hypothesis tests, it finds evidence for dependence in both directions.
- Effects of Author Centrality on Citation Success: Authors of highly successful papers are more central in the coauthorship network than authors of less successful papers across the tested centrality metrics.The distributions are compared using one-sided Wilcoxon-Mann-Whitney tests, with p-values below 0.01 for all four metrics.
- Effects of Author Centrality on Citation Success: The study defines a paper’s coauthorship centrality as the network centrality of its highest-degree coauthor.This operationalization is motivated by the assumption that the best-connected coauthor contributes the largest socially triggered visibility.
- Effects of Author Centrality on Citation Success: For Top 10% papers and authors, k-core centrality yields P(topmetric|toppaper) = 0.21 and P(toppaper|topmetric) = 0.22.These probabilities are below 1, indicating that centrality does not provide a simple linear correspondence with citation success.
- Effects of Author Centrality on Citation Success: Considering the intersection of all four centrality metrics, P(toppaper|topmetric) = 0.36 for the Top 10% group.The authors use this observation as the basis for a naive Bayes classifier.
- Coevolution of Coauthorship and Citation Success: Authors whose citation success increases show significant increases in k-core, betweenness, and degree centrality, while declining-success authors show significant drops in k-core, eigenvector, and degree centrality.The results support both hypothesized directions of coevolution between citation success and future coauthorship centrality.
- Coevolution of Coauthorship and Citation Success: A highlighted author moved from the coauthorship-network periphery to its center over five years, with increased degree and betweenness centrality.The example concerns an author who lacked a highly successful paper in 2002 but had one in 2007.
5 Predicting Successful Publications
The paper tests whether coauthorship-network centrality can predict whether a publication will become highly cited five years later. It compares simple centrality-based rules with a multidimensional Random Forest classifier using author-position features.
- A naive Bayes rule requiring top-10% betweenness, degree, k-core, and eigenvector centrality achieves precision = 0.36% and recall = 0.15%.The rule predicts top-10% citation status when an author is simultaneously in the top 10% for all four centrality measures.
- Each publication is represented by 10 features comprising maximum and minimum degree, eigenvector, betweenness, and k-core centralities, plus the number of coauthors.The features are computed from the coauthorship network in the window [t-2,t].
- The Random Forest classifier achieves 60% precision, identifying top-10% papers six times more effectively than random guessing.Random guessing has 10% precision in this setting.
- The classifier achieves 18% recall, an 80% improvement over the 10% recall expected from random guessing.It correctly identifies about one fifth of all top-10% papers in the field.
- The results indicate that citation success depends on multiple coauthorship-centrality measures rather than a single centrality value.The Random Forest outperforms the naive Bayes classifier that combines thresholded centrality conditions.
6 Discussion and Conclusions
The paper finds a strong relation between authors’ positions in coauthorship networks and their future citation success, while emphasizing that citation outcomes are not shown to be causally determined by centrality. Its conclusions are limited by data quality and the focus on computer science, but support using multiple centrality measures to study the social dimensions of citation behavior.
- Limitations: The analysis is constrained by the completeness and correctness of its data and by its restriction to computer science.The authors note that publication practices vary across disciplines, limiting generalization to other research communities.
- Limitations: The observed statistical dependence does not establish a direct causal relation because centrality and citations could both reflect researchers’ scientific excellence.The authors explicitly do not rule out such alternative explanations.
- Contributions: No single investigated centrality measure accurately predicts an author’s future citation success; combining complementary measures is crucial.The measures capture different dimensions of network importance, including visibility and social influence.
- Contributions: A Random Forest classifier using multidimensional coauthorship-centrality features predicts future citation success with remarkably high precision.The prediction uses only the social embedding of the authors at publication time.
- Contributions: The study provides evidence of a strong relation between authors’ collaboration-network positions and their future citation success.It presents a large-scale analysis using time-evolving collaboration networks and citation numbers.
- Conclusions: The findings are intended to clarify citation dynamics and inform debate about the meaningfulness and use of citation-based evaluation measures.The authors caution against viewing automated success prediction as their primary contribution because widespread adoption could affect scientific culture.
Supplementary Material
The supplementary material defines network-centrality and classifier-evaluation measures, then describes Random Forest prediction of citation success. It also reports that citation counts five years after publication show no significant correlation with authors’ centrality metrics.
- Network centrality measures: Degree centrality counts an author’s direct coauthorship neighbors and is a local network measure.Because the coauthorship network is undirected, degree centrality is simply the number of direct neighbors.
- Network centrality measures: Eigenvector centrality is high when an author connects to other authors with high eigenvector centrality, capturing inherited importance beyond neighbor count.
- Network centrality measures: K-core centrality measures how deeply an author is embedded in the network, requiring persistence through successive removals of low-degree nodes.It is not additive: many low-k-core neighbors do not necessarily compensate for few high-k-core neighbors.
- Correlation analysis: Citation numbers five years after publication showed no significant Pearson or Spearman correlation with authors’ social-network centrality metrics.The supplementary text states that none of the reported correlation results supports a significant relationship.
- Classification: Random Forest classification grows multiple decision trees on bootstrap samples and assigns each case the class receiving the most tree votes.The method selects a random subset of variables at each tree split and averages across the resulting forest.