| Abstrakt: |
• Interpretable machine learning approaches are applied to the task of event-based Group Evolution Prediction of scientific communities. • Interpreting the prediction models of community evolution can reflect the underlying mechanisms of the evolution of scientific community. • A detailed feature set, including topological, external, core node, and temporal attributes, is compiled. • Shapley Additive exPlanations (SHAP) is applied to interpret the best model, which quantifies the contributions of features. Scientific communities serve as a fundamental structure of academic activity, and its evolutionary behavior also reveals the development of science. To track the evolution of scientific communities and dig into the mechanism behind it, we formulate the task of event-based Group Evolution Prediction and apply interpretable machine learning approaches to the task. Seven evolution events for prediction are defined based on the evolution chains of scientific communities detected from the collaboration network. By using a detailed feature set, including topological, external, core node, and temporal attributes, Extreme Gradient Boosting, and Random Forest are adopted for the prediction models. Experiments on the dataset of Library and Information Science shows that Random Forest performs the best, with the F1 scores of five events greater than 0.60. Shapley Additive exPlanations measure is applied to interpret the best model, i.e., quantify the contributions of features. It is observed that connectivity within a community has the most crucial influence, and community size, research topic consistency, research topic diversity, average node age, and the ratio of intermediary nodes play vital roles. The proposed methodology offers a solution to unearth the underlying mechanisms of the evolution of scientific communities, and the findings could be useful for scholars and policymakers to monitor scientific communities and take proactive actions. [ABSTRACT FROM AUTHOR] |
Full Text from ScienceDirect