Prediction and interpretation of comprehensive performance of large-scale instruments in universities using interpretable machine learning

To promote data-driven management of large-scale scientific instruments in universities, this study builds a predictive model for the instruments’ comprehensive performance score (CPS) and identifies its core influencing factors by integrating machine learning (ML) and interpretable analysis. Based on operational data of 223 large-scale instruments across 18 departments of a university, seven core indicators were selected as input features. Nine models, including three linear baselines (Linear, Ridge, and Lasso) and six classical ML algorithms (KNN, SVM, DT, RF, XGBoost, CATBoost) were compared using R 2 , MAE, RMSE and MAPE, with descriptive statistics and Pearson correlation analysis for data preprocessing, and SHAP method for feature importance quantification and impact pattern analysis. Results show that ensemble learning models outperform traditional ones, with XGBoost achieving optimal test-set performance and excellent generalization, showing no signs of overfitting or underfitting. SHAP analysis reveals that annual effective external operating hours (35.3%), internal income (27.9%) and sharing rate (20.3%) are the core determinants, contributing over 80% of the model’s predictive power and exerting significant positive monotonic impacts. Machine hour utilization rate has a moderate positive effect, original equipment value shows a complex non-linear impact, and external income and annual newly added functions have negligible influence. This study provides evidence for XGBoost’s superiority in instrument performance prediction, and clarifies that extending external service hours, increasing internal income and improving resource sharing are the factors most strongly associated with higher instrument performance in our model. The findings provide a scientific basis for universities to optimize instrument resource allocation and enable the shift of instrument management from experience-based to data-driven practice, with important implications for the intelligent management of large-scale scientific instruments in higher education.