Decoding the Impact Dynamics: Prediction and Interpretation of Scientific Research Cited by Policy Documents

Purpose The translation of science into policy is frequently oversimplified by linear assumptions, leaving the specific micro-level drivers of policy adoption unclear. This study aims to develop a systematic and explainable framework to predict and interpret how structural, academic, and dissemination factors influence the policy citation of scientific papers. Design/methodology/approach This study constructs a large-scale dataset linking scientific research to policy documents, and evaluates multiple machine learning classifiers to predict policy citations, employing SMOTEENN to address class imbalance. To uncover the decision-making mechanisms, the SHAP framework is applied to the XGBoost model. This enables an in-depth analysis of global feature importance, contribution directions, and non-linear dependencies. Findings The results show that policy citation is moderately predictable and is better captured by ensemble-based machine learning models than by simpler baselines. SHAP analysis reveals a differentiated and non-linear selection process. Citation Count is the strongest predictor, while Patent Count and Tweet Count also provide substantial but distinct predictive information. The dependence plots further identify saturation effects, power-law-like decay, threshold effects, and bounded novelty. These findings support a Validation–Knowledge Filtering Mechanism, in which scientific papers are more likely to be cited in policy publications when they are both externally validated and suitable for policy-oriented knowledge use. Research limitations The focus on articles from science and Overton-indexed policy documents may limit the generalizability of the findings. Additionally, the binary classification of citations overlooks the substantive degree and contextual depth of the extent and context of how scientific evidence is used in policy documents. Practical implications Enhancing policy impact requires more than academic impact, institutional prestige, or media exposure alone. Researchers should strengthen both evidence credibility and policy-oriented usability, while policy organizations should avoid over-reliance on simple prestige, scale, or visibility signals. Originality/value This study advances scientometrics by shifting from descriptive correlations to mechanistic prediction. It identifies the drivers of policy citation and proposes a Validation–Knowledge Filtering Mechanism for explaining scientific paper selection in policy.