Hybrid prediction models for assessing the Higher Education Institutions Performance in QS World Institution Rankings

1. Introduction

Higher education institutions, such as universities and colleges, are assessed and compared using lists or other methods that consider a variety of criteria and aspects. These rankings are invaluable resources to help prospective students and their families make well-informed decisions about where to pursue higher education. Rankings help students find colleges that best fit their academic and professional objectives by offering insights into the schools’ general caliber, standing, and effectiveness. Additionally, they can assist establishments in identifying areas in need of development and formulating plans to elevate their profile in the international higher education arena. University rankings are vital for several reasons and benefit many stakeholders in the higher education system. First, they help students and their families make well-informed decisions by offering a prompt evaluation of the caliber and standing of an institution and by assisting in the selection of colleges that complement academic and professional objectives. Rankings also function as a type of quality control, highlighting that universities continually provide resources, research, and instruction of the highest caliber. Rankings assist scholars in identifying colleges that are well-known in particular subjects, which facilitates the search for partners and specialized resources.

Furthermore, elite professors, researchers, and students are drawn to highly regarded universities to enhance the learning environment and promote intellectual diversity. Internationally recognized rankings offer global awareness and can help improve an institution’s reputation and appeal to donors, which can aid fundraising efforts. For universities, self-evaluation and benchmarking are essential tools that support their efforts to improve quality and pinpoint opportunities for development. Rankings help government agencies and legislators make decisions on how to allocate resources and advance transparency and public accountability in higher education. Owing to its ability to analyze vast datasets, discover patterns, and make informed forecasts, machine learning plays a critical role in prediction across a wide range of domains, including forecasting stock prices, currency exchange rates, market trends, assisting investors and traders, recommendation systems, autonomous vehicle decision-making, medical condition diagnosis, outcome prediction, and disease risk assessment. Its use is widespread and growing. Machine learning plays an important role in predicting university rankings, because it allows institutions to combine historical and various data sources to create more accurate forecasts of their future rankings. These forecasts enable institutions to take proactive steps to improve their academic and research skills, faculty qualifications, and overall appeal to students and researchers, thereby increasing their global ranking.

In this study, we used machine learning methods to predict QS World University rankings. We start with cleaned, preprocessed data and divide them into two parts: 70% for training and 30% for testing. To construct a prediction model, we applied two different groups of algorithms such as Hybrid machine learning optimization algorithms, including i) Ridge Regression, Long Short-Term Memory (LSTM) and LightGBM, ii) Particle Swarm Optimization (PSO) and Tabu search (TS) with base models, and iii) PSO and TS with hybrid models. Hybrid machine learning non-optimization algorithms include i) Gradient Boosting (GB) and k-nearest neighbors (KNN); ii) Random Forest (RF) and XGBoost (XGB), Support Vector Machine (SVM), Neural Network (NN), and GBM; and iv) NN and XGB. Metrics including the Coefficient of Determination (R-squared) Score, Mean Square Error (MSE), Root Mean Square Error (RMSE), and Accuracy Percentage (%) were then used to compare the outcomes. Finally, based on these metrics, we determined which model is more effective for predicting university rankings.

2. Literature study

University rankings have evolved as a significant means of measuring the quality of higher education institutions in response to global demands and the demand for transparency and efficiency in public organizations. In Ref. 1, Stefan Wilbers and Jelena Brankovic delved into the historical-sociological pillars of university rankings, concentrated in the United States. This study highlights the shift in the perception of organizational success that occurred during the postwar period, a shift that was impacted by functionalism’s ascent as the prevailing theory. An indication of this shift is the grading system for graduate programs, which was established by the ACE, the American Council on Education, and the National Science Foundation between 1966 and 1970. In the 1970s, social scientists began to examine the criteria and utility of categorizing institutions of higher learning, and they still do so now. Dembereldorj² assessed the impact of worldwide university rankings on institutions around the world, underlining their significance in both developed and developing countries. While universities in developing economies place more emphasis on research intensity, institutions in advanced economies prioritize competitive competency to achieve top rankings. This study makes the case that resource constraints and the demand for institutional or competitive competency drive worldwide rankings, which in turn actively affect higher education.

Vernon et al.³ insist on the necessity of making real efforts to raise the caliber of research, focusing on developing novel standards that organizations can use to evaluate and enhance their social contributions. Prioritizing quality above quantity is essential for confirming efforts to boost research output, which leads to advancements in science, economic expansion, and public health. The study highlights the deficiencies of current standards and recommends evaluating research outcomes in three dimensions–scientific influence, economic effects, and public health consequences–to provide a comprehensive assessment of research performance within the context of an academic institution. In Ref. 4, Leah Dowsett focused on how Australian universities developed long-term strategies for worldwide rankings over fifteen years by assessing four colleges and discovered that these rankings greatly influenced their research programs. Institutions’ market positioning and rankings have improved as a result of their proactive engagement with the rankings and efforts to influence and react to them. Murat Perit C¸ akır et al. compared and contrasted worldwide and national university ranking systems⁵ to highlight divergent views. Global rankings concentrate on research achievement with fewer parameters, whereas national rankings include a wider range of indicators such as institutional and educational components. The data show little association between country and global rankings, with a few exceptions. In some countries such as Brazil, Chile, and Poland, national rankings are predicted by global rankings, suggesting a relationship between research success and educational criteria. Owing to the challenge of acquiring accurate per capita data for global rankings, bibliometric criteria related to size have gained increasing attention. National rankings are becoming increasingly prevalent, especially in underdeveloped nations. This creates opportunities for benchmarking and comparative studies that can enhance international ranking systems and provide information on higher education.

In Ref. 6, Adina-Petruta Pavel compares and contrasts the methodology, standards, and effects on stakeholders of three influential worldwide university rankings: Quacquarelli Symonds (QS), The Times Higher Education World University Rankings, and the Academic Ranking of World Universities (ARWU), commonly referred to as the Shanghai Ranking. Notably, the global rankings place research ahead of instruction. The findings inspire institutions to improve their operations and place greater emphasis on industry collaboration and innovation. Friso Selten et al. analyzed the techniques of the three significant global university rankings stated above in Ref. 7 and discovered that, while these rankings are somewhat stable over time, there are differences between them. Applying factor analysis with principal component analysis, the study indicates that the differences in these rankings fundamentally evaluate two factors: university prestige and research performance. By correlating and visualizing these data, it is possible to discern between ranks. This paper responds to common criticisms of the ranking technique by stating that the variables may not accurately represent the concepts they are supposed to represent. It emphasizes how difficult and imprecise it is to use rankings to judge a university’s progress.

In Ref. 8, Moskovkin et al. explored the global battle for university reputation, which began in 2003, focusing on quantitative comparisons of World University Rankings by ARWU, QS, and THE from 2014 to 2018. The study discusses the remaining challenges with ranking consistency, as well as the approach of integrating the number of institutions and Overall Scores by nation. This study examines differences in university rankings to identify the top 100 rankings that are most and least consistent within the top 100 rankings. It also delivers the most favorable aggregated indicator values for the US and United Kingdom. Shehatta and Mahmood collected, examined, and analyzed the top 100 institutions qualitatively and quantitatively according to six significant global rankings published in 2015.⁹ The six global rankings that were selected were the Academic Ranking of World Universities (ARWU), Quacquarelli Symonds World University Rankings (QS), Times Higher Education World University Rankings (THE), National Taiwan University Ranking (NTU), US News & World Report Best Global University Rankings (USNWR), and University Ranking by Academic Performance (URAP). For comparison, the number of overlapping universities and the Pearson’s and Spearman’s correlation coefficients between every pair of the six global rankings under investigation were used. They found similar traits or independent factors used to determine rank in all the ranking systems studied. Bublyk et al. explored the elements impacting global university rankings in Ref.10 using data from the QS World University Rankings. It employs statistical and correlation analyses and focuses on how the ranking of Lviv Polytechnic National University has changed over time. This research identifies trends, benefits, and drawbacks and establishes a framework for worldwide university growth plans. It also assesses information security compliance, and provides comprehensive strategic recommendations for future progress.

Hasan and Abuelrub¹¹ compared the usability of Jordan’s top three university websites on the outcomes of the heuristic assessment method and Eduroute, one of the main ranking systems. The findings suggest that information regarding the general usability of university websites can be obtained from Eduroute’s rankings. The heuristic evaluation technique used in this study also revealed common usability problems found on university websites. Mahesh provided an overview of several machine learning techniques that might be applied to predict and categorize the data provided in Ref. 12, which serves as a foundation for subsequent predictions of university rankings. Liu et al.¹³ examined the institutions and their rankings in each country and discovered a correlation between six indicators: academic credibility, employer credibility, staff/student ratio, citations per staff, global staff ratio, and global student ratio. ML algorithms such as There are three methods used: XGBoost, random forest, and linear regression. The mean absolute error, mean squared error, and root mean squared error were used to gauge the accuracy of the models. The results demonstrate that XGBoost has lower-precision measurement values. The lower the value, the more accurate the prediction. In Ref. 14, Gadi Himaja et al. compared and contrasted various regression techniques to recommend a rank prediction system to a national institute. They identified the most accurate one using evaluation metrics such as R2, MAE, MSE, and RMSE, and they chose the Random Forest using threshold value comparison and z-score calculation.

In Ref. 15, Vaibhav Singh et al. used a standardized database rating of the world’s universities by Times Education. The dataset was split into test data from 2016 and training data from 2011 to 2015. Using linear regression, they calculated the expected rank score for teaching, research, citations, and international orientation. Finally, they used the expected total rank score to rank the universities globally. Tabassum et al.¹⁶ developed a global university rank prediction algorithm using The Times Higher Education World University Ranking, which was established in the United Kingdom in 2010. Data analysis of prior university rankings by country was conducted to discover the most influential elements or indications for prediction. Outlier detection algorithms were utilized to generate predicted scores, whereas the rank_score_calculate algorithm was used to predict the feature scores. Universities have been ranked worldwide. Methods for evaluating the suggested model: The number of matched ranks vs. rank deviation, recall, ROC curve, and accuracy vs. deviation are all variables to consider. Li¹⁷ discovered that utilizing linear regression to evaluate various indicators can better predict the comprehensive score of colleges based on many features. They used a radar graphic to compare the indicators they looked at. Following the investigation, faculty quality was found to be a key determinant of institution ranking. In Ref. 18, Dr. Prakash Kumar Udupi et al. created machine learning models to anticipate global rankings and examined the Quacquarelli Symonds (QS) method of analyzing university rankings worldwide. This study analyzes the information using exploratory data analysis and then evaluates machine learning algorithms using regression approaches to predict worldwide rankings. The QS system rankings are divided into three categories: worldwide overall rating, regional, and global ranking according to the subjects. Key performance metrics, including teaching, research, employability, university mission, and internationalization, form the foundation of QS global rankings. Boosting regression makes use of the Gaussian loss function to improve the predicted outcomes. The same evaluation metrics as those in Ref. 13 were used.

In Ref. 19, Estrada-Real et al. highlighted that we undertook a Feature Selection exercise to assess the validity of the six indicators using the Recursive Feature Elimination (RFE) method, with the rationale for using the QS technique with the availability of data on their website for ranking purposes. Using supervised machine learning techniques, including multiple regression with panel data, logistic regression, decision trees, random forests, and support vector machines, they created a prediction model with categorical response variables based on the generated training data. Test data and statistical measurements, including R2, p-value, accuracy, sensitivity, specificity, confusion matrix, receiver operating characteristic (ROC), and Area Under the Curve (AUC), were used to evaluate the model’s output. Both logistic regression and random forest analyses yielded comparable results. In Ref. 20, Yan-yan SONG and Ying LU investigated how decision trees are commonly employed in data mining to construct classification systems and forecasting models. In Ref. 21, Nishi Doshi et al. used exploratory data analysis (EDA) to analyze ranking data using correlation heatmaps and box plots, and introduced a novel method for extracting decision paths for rank improvement using Decision Tree (DT) algorithms and data visualization. The approach, which has been refined with Laplace correction, provides institutions with a quantitative means to analyze development potential, plan long-term actions, and design distinctive success roadmaps. Sziklai applied The Weighted Top Candidate (WTC) approach in Ref. 22 to rank smaller academic institutions in specific research areas. Jajo and Harrison provided an overview of the available ranking systems in Ref. 23, including the founder, year, range, and measures. They employed partial least squares path modelling to create an achievement index that can be used to compare university performance across several ranking systems simultaneously. Roba Elbawab utilized unsupervised machine learning to group institutions.²⁴ The results divided the universities into four categories. Mohamed El Mohadab et al. evaluated a variety of classification strategies,²⁵ including supervised, semi-supervised, and unsupervised learning. The study used performance metrics, such as F-Measure, Precision, GMAP, NDCG, MAP, and Recall, to determine relevant characteristics within the research setting.

Furthermore, the ensembled models constructed using SVM, KNN, MLP, decision trees, random forest and Logistic regression algorithms with fast Fourier transformers were used in Ref. 26 for predicting international ranks of HEIs utilizing the Shanghai world ranking dataset. Wardley et al, have used different machine learning models to rank universities in Canada on different categories and recommended gradient boosting as the best performer.²⁷ Also, the predictive power of the ensembled models was further improved by trimming the low-performing model and optimized through heuristics, in predicting international rankings.²⁸ The power of machine learning models was further leveraged for predicting research performance indicators,²⁹ job satisfaction analysis³⁰ and financial stability planning³¹ in HEIs. Few attempts to rank HEIs using multicriteria decision-making models with machine learning and NLP were carried out.^32–35 These models were constructed based on subjective and objective measures to strengthen the results.

Our survey of research papers culminated in a systematic tabulation and is tabulated in Table 1 (Extended data), capturing essential aspects including Dataset Diversity, Ranking Level, Number of Algorithms employed, Highest Accuracy attained, identification of the Best Algorithm, and an insightful overview of limitations. This structured presentation offers a concise and comprehensive snapshot of the various methodologies and outcomes prevalent in the landscape of research endeavors.

1.4.1 Methods

Figure 1 shows the proposed methodology.

Figure 1. Proposed architecture.

Figure 2. The used dataset's correlation heat map.

Figure 3. Conversion of categorical values to numerical values.

Figure 4. The scatter plot for the hybrid approach.

Figure 5. Scatter plot of the features.

Figure 6. a. Comparison of performance of the hybrid algorithms Model 1 to Model 3. b. Comparison of performance of the hybrid algorithms Model 4 to Model 7.

Figure 7. a. Comparison of individual metrics of the hybrid algorithms HRRM, LSTM, LightGBM and PSO. b. Comparison of combined visualisation of all metrics for hybrid algorithms HRRM, LSTM, LightGBM and PSO.

Figure 8. (a). RMSE contrasts the algorithms under consideration. (b). MSE comparison of the considered algorithms. (c). R-squared comparison of the considered algorithms. (d). Accurate comparison of the algorithms under consideration. (e). Recall comparison of the algorithms considered. (f ). Comparison of the F1 Scores of the evaluated algorithms. (g). Accuracy comparison of the considered algorithms. (h). Comparison of the overall performance of selected algorithms.

3. Results & Discussions

The dataset used in this study stems from the esteemed QS World University Rankings for 2024, a comprehensive assessment comprising 1,500 institutions across 104 global locations. With 1,499 rows and 29 columns, it encompasses vital attributes, such as 2024 and 2023 ranks, institution details, size, focus, research metrics, reputation scores, and newly introduced metrics for sustainability, employment outcomes, and international research networks. These additions reflect a methodological evolution, enabling a nuanced evaluation of universities’ commitment to social responsibility, employability, and global research collaboration. Notably, the dataset provides a holistic view, including academic reputation, employer perception, faculty-student ratios, citation impact, and internationalization indices. Widely recognized and respected, the QS World University Rankings influence global perspectives on higher education, making this dataset a valuable resource for researchers, policymakers, academics, and prospective students, offering multifaceted insights for education and higher education policy analysis. Because there are no irrelevant characteristics in the dataset, it contains correlated features, as shown in the correlation matrix in Figure 2.

In our initial research phase, we worked with raw data, a dataset that included various college parameters including reputation ratings and rankings. Missing values presented analytical difficulties and required correction, which is typical for raw data. Categorical data, such as institution names, were numerically encoded for machine learning model compatibility, and scaling techniques were used to handle the different scales of numerical variables (see Figure 3). Mitigating possible influences on model performance requires the discovery and rectification of defects or outliers. After thorough preparation, we performed a transformational refinement of our dataset, which resulted in decreased errors, controlled outliers, and transformed categorical variables. The numerical feature scale homogeneity was guaranteed by standardization and normalization processes. The dataset was enhanced with new characteristics that came via mathematical processes, making it ready for a variety of analytical uses such as machine learning. Dimensionality reduction techniques were applied, which streamlined the dataset without compromising important information and accelerated the modelling process. Our dataset is best suited for model training and assessment after pre-processing, providing improved interoperability with different machine learning techniques.

Initially, we preprocessed the dataset in the given model generation procedure. This includes addressing missing data, encoding categorical variables, scaling numerical features, and generating newly derived characteristics. Next, the dataset was divided into the testing and training sets. We built hybrid prediction models that use optimization without optimization parameters. To evaluate the prediction performance of both models, we computed several evaluation measures, such as the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R²) score. Interestingly, we also added a statistic called “accuracy percentage” to measure how accurate ranking predictions are within a given threshold. The results are presented in Tables 3 and 4. The results were synthesized using a comprehensive bar plot. A further simulation is shown in Figure 6 and the execution using the hyperparameter configurations is presented in Table 2. The proposed approach performed better in terms of convergence than other algorithms. Regarding the total RMSE, MSE, and R² Score the performance of the proposed method in comparison with the other algorithms is displayed in Figure 9.

Figure 9. (a) Hybrid Model – 1. (b) Hybrid Model – 2. (c) Hybrid Model – 4. (d) Hybrid Model – 5. (e) Hybrid Model – 6. (f ) Hybrid Model – 7.

4. Conclusions and future work

Finally, by proposing and analyzing hybrid algorithms, we hope to improve the accuracy of predicting QS World University Rankings. With the lowest RMSE, highest R2 score, and commendable accuracy, Hybrid Model 3 emerged as a top performer using Particle Swarm Optimization and tabu search. Furthermore, Hybrid Model 5, which combines Gradient Boosting and k-nearest neighbors, outperformed the competition. Expanding deep learning techniques, such as tailored neural network designs and utilizing pretrained language models, are future directions for ranking prediction. Integrating multimodal data, stressing explainability, and addressing ethical problems such as fairness and transparency are critical.

Recognizing limitations, such as dataset specificity and the necessity for external validation, researchers should investigate qualitative characteristics while keeping in mind the changing landscape of higher education. Ethical issues, data quality improvements, and continual deep-learning breakthroughs are critical for refining ranking forecasts and ensuring their relevance and applicability in the changing sector of university assessments.