A Hybrid Framework for Predicting Citation Harm from Retracted Papers

The increasing prevalence of retractions poses significant challenges to academic integrity and scientific progress. While various tools track retracted papers, the academic community lacks proactive mechanisms to predict the potential harm caused by citing retracted research. Building upon recent work quantifying the harm, we construct a dataset to forecast the harm papers will suffer from citing retracted research. We further propose a hybrid framework that combines language model with gradient boosted trees. Our framework achieves the lowest mean absolute error (MAE) across all prediction time horizons, achieving a 2.4% reduction in MAE compared to baseline methods. This work provides foundational resources and methodological insights for future studies on retraction impacts, aiming to help researchers make more informed citation decisions.