Hybrid machine learning framework for forecasting and evaluating university rankings

The growing importance of university rankings in higher education requires evaluation methods that are both predictive and interpretable. Traditional ranking approaches are usually based on fixed weighted indicators, while purely data-driven machine learning models often provide limited transparency for institutional decision-making. This paper proposes a hybrid analytical–machine learning framework for evaluating and forecasting university rankings using multidimensional performance indicators. The framework combines an interpretable weighted aggregation model, an XGBoost-based residual correction mechanism, an LSTM forecasting module, and SHAP-based explainability analysis. The empirical dataset was constructed as a university-year panel covering 20 Kazakhstani universities over the 2019–2024 period, resulting in 120 observations. The input indicators included publication output, citation impact, h-index, international collaboration, and research funding. The models were evaluated using a temporal validation protocol, with 2019–2022 used for training, 2023 for validation, and 2024 as the independent test subset. The results show that the hybrid model reduced RMSE from 0.185 to 0.092 compared with the analytical model and from 0.124 to 0.092 compared with the standalone XGBoost model. Similar improvements were observed for MAE, MAPE, and ranking error. However, the remaining deviations indicate that the model should be interpreted as reducing, rather than eliminating, prediction uncertainty. SHAP analysis showed that citation impact and publication output had the widest feature-contribution distributions, while all indicators demonstrated both positive and negative effects depending on the observation. The proposed framework can support institutional analytics by combining ranking evaluation, forecasting, and interpretable identification of key performance drivers.