Bibliometric feature identification and analysis of retracted papers in biomedicine: An interpretable machine learning perspective

Abstract

Nowadays, paper retraction is becoming a serious problem in academia, particularly in biomedicine. Previous studies in this area have examined various features of retracted papers. Based on these findings, the aim of this paper is to construct a model to predict potential retraction cases, and analyze the retraction features over time and across fields. Specifically, we construct an XGBoost model using 9424 normal and retracted biomedical papers published between the year 1983 and the year 2023 from the Web of Science Core Collection database. This model has an accuracy of 87 %. Nine important features are identified, ranked, and their contributions to the model are discussed using interpretable machine learning techniques. Moreover, heterogeneity analysis by publication year reveals that the importance of these features has changed over time. The generalizability of the model is validated in the field of computer science (98.12 %) and telecommunication (74.92 %). Finally, we analyze the similarities/differences in these features among the three fields. The result of this study confirms the features identified by previous studies. Further, the way that these features describe and predict whether a paper is retracted or not is revealed by interpretable machine learning techniques. This has not been discussed much in previous studies. Additionally, this study provides details on how these features change over time and across disciplines in predicting retractions. Finally, the results of this study may shed some new light on further research. It may also be used as a reference in science policy-making.