Study on the predictability of new topics of scholars: A machine learning-based approach using knowledge networks

Abstract

Scholars continuously explore new research topics to drive personal academic achievements. While factors influencing topic selection exist, the predictability of scholars’ choices regarding new topics is not yet fully understood. To bridge the gap, this study investigates the predictability of new topics of scholars (NTS). The research task is transformed into a binary classification, predicting whether NTS that appear in the disciplinary knowledge network will be adopted by a scholar in the future. Using PubMed Knowledge Graph (PKG) as the data source, over 17,000 local knowledge networks (LKNs) of individual scholars are constructed, along with a global knowledge network (GKN) of all the scholars in the database. Sixteen features of knowledge network topology and candidate topics are extracted, and seven machine learning algorithms are applied. Our large-scale experiments show that the best prediction model achieves an F1 score of 86.49%. Shapley values provide more interpretable results. A 1-year observation window appears to be sufficient for making predictions. Novel topics and young scholars exhibit good predictability. Our findings provide profound insights into the predictability of scholars' topic selection and offer practical implications for future in-depth studies.