Beyond Citations: Explainable Machine Learning for Identifying Early-Career Researchers from Bibliometric Data

Abstract Identifying ECRs (early career stage researchers) is crucial for developing academic collaborations and supporting emerging scholars, but in the context of large bibliographic datasets it is a challenging task. In this paper, a machine learning framework for automated identification of ECRs is proposed, with a particular focus on distinguishing between early career stage researchers and more established researchers. Citation based and non-citation based features are leveraged in the creation of a training set, which consists of publication and collaboration related features extracted from the AMiner dataset with additional annotations to represent ECR researchers. We evaluate different feature selection and extraction techniques as well as several supervised classification algorithms including Random Forest, Logistic Regression, Support Vector Machines, XGBoost and Gradient Boosting, finding that the best performing models (XGBoost and Gradient Boosting) achieve 99% accuracy in ECR classification. The application of explainable AI methodology provides further insights into the contribution of different features to the classification task, revealing that the most informative non-citation features are related to trajectory based analysis. Moreover, the analysis shows that non-citation features have greater discriminatory power for ECR identification compared to citation features, illustrated in the OpenAlex case study. The proposed approach enables more accurate recommendation of potential collaborators as well as more informed policy design related to the needs of the ECR community.