Predicting scientific breakthroughs based on knowledge structure variations

Abstract

Breakthrough research plays an essential role in the advancement of the scientific system. The identification and recognition of scientific breakthroughs is thus of extreme importance. We propose a citing-structure perspective for observing the unfolding of breakthrough research from variations in knowledge structure. The hypothesis is empirically validated that scientific breakthroughs show distinctive knowledge structure characteristics, which are further utilized to predict breakthroughs in their early stage of formation. These characteristics include average clustering coefficient, average degree, maximum closeness centrality, and maximum eigenvector centrality in the direct citing networks of a breakthrough publication. Several explanations are provided for the effectiveness of the predictive models. We also show that: (1) the number of direct citation counts is of low predictive power, with even a negative impact on prediction performance; (2) disciplinary differences exist in knowledge structure, and this should be taken into account; (3) breakthrough characteristics are most prominent in the first layer of citing networks; (4) timing is critical, and 2- to 3-year-old citing networks have greater predictive power.