A Comparative Benchmark of Machine Learning Models for Predicting University Rankings: Insights for Higher Education Analytics

Abstract

QS World University Rankings are so important for evaluating academic performance and guiding strategic decision-making. Though great strides have been made to address this using machine learning and deep learning techniques, the majority of prior works suffer from outdated data use, lack of complete comparisons between state-of-the-art algorithms, and a lack of feature selection procedures that can cause bias in results beneficial towards a specific algorithm. This paper aims to fill these gaps and provide an integrated methodological approach. The analysis was based on the latest QS data (2023 and 2024) from 1,498 institutions and 26 performance indicators. After the data cleansing process, a multi-criteria weighting procedure (Information Gain, Correlation Coefficient, Relief algorithm and Gini Index) was used to evaluate feature importance with the aim of only selecting attributes inside each cross-validation training loop based on heuristic rules (the PART approach) so as not to be faced with performance penalties. Next, a total of six different predictive models were constructed and tested: (1) Linear Regression; (2) Decision Trees; (3) Random Forests; (4) Gradient Boosting Trees; (5) Multi-Layer Perceptrons and finally, (6), Convolutional Neural Network. The overall results showed that both Gradient Boosting Trees and Convolutional Neural Network models achieved the best performance among the evaluated models under the current experimental setup with the highest R-squared (0.98) together with their predictive accuracy 97.8% and lowest Mean Squared Error of 0.06 respectively. Feature importance analysis indicated that Academic Reputation and Research Impact are the most important features. This study adds value by: (1) proposing a transparent and reproducible approach that incorporates extensive benchmarking with conservative feature selection, (2) updating the performance benchmark through recent data and (3) providing meaningful insights for evidence-based strategic decision-making in higher education institutions. However, the findings are limited to QS datasets (2023–2024) and may not generalize to other ranking systems or contexts.