**Table 7.** Performance of rotation forest through various feature selection techniques.

### **5. Discussion**

In the past, most educational research has been focused on and evaluated students' academic performance in specific institutions or regions. Performance was calculated with consideration for various influences including socio-economic and demographic factors as well as students' personal, family, and academic backgrounds. Apart from student academic results, other factors also have a substantial impact on the performance of any educational institution. The present study focused on the importance of other highly influential factors along with student academic results, such as the students per teacher ratio, the number of schools in a region, whether schools were located in rural or urban areas, the availability or lack of classrooms, electrical facilities in schools, availability or lack of furniture for students, open-air classes, computer lab facilities, science labs, and playgrounds in schools. Previous research [44–48] suggested that data pre-processing (normalisation, discretisation) techniques enhanced classifier performance, as these techniques reduce the biases among features. Furthermore, related studies showed that the min-max normalisation method performed better than other data normalisation methods [49–51]. It has also been observed in related studies that binning-based data discretisation techniques outperformed other techniques based on their results [52–54]. This study will help in the identification of underperforming regions based on institutional performance. It will also support governance in performance monitoring, policy formulation, target-setting, evaluation, and reforms to address the issues and challenges of education. In this research, various feature selection methods were combined with machine learning models to obtain efficient results. The fourteen most significant features were used, as selected through feature selection methods. The J48 decision tree model was combined with feature selection methods. The best results were obtained using the relief-F-based feature selection technique, which achieved maximum accuracy of 68.5% with an ROC value of 0.63. The highest accuracy of 68.5% was achieved with the SVM (RBF kernel) model while employing the relief-F based feature selection method. After the SVM model, the random forest was utilised to obtain more effective results. The highest accuracy of 71.3% with an ROC of 0.65 was obtained. After the random forest, the rotation forest was utilised to obtain more effective results. The highest accuracy of 73.2% was obtained. After J48, SVM, random forest, and rotation forest, artificial neural networks were used to achieve more efficient results. It has been observed that this model predicted the best results with an accuracy of 82.9% while utilising the relief-F based feature selection method. The artificial neural network outperformed and yielded the highest accuracy, of 82.9%, among the five classifiers employed in this study. The performance of ANN also proved efficient while evaluating other performance metrics. It was also observed that the target class (medium) results were better than other target classes (low and high). This is because the number of instances in the medium class were significantly higher than in the high and low classes. The performance of machine learning models is better when trained on large datasets. In our study, the performance of machine learning models on medium classes was also high due to the large amount of data as compared to other classes.

This study provides additional support for researchers to employ the ANN model and apply it to social science studies. Moreover, this study showed that there is value in including special education-related predictors to improve classification accuracy. The study demonstrated how geographical and demographic variables could all add to the classification accuracy of prediction models. Lastly, the study results offered strong evidence that school facilities are highly predictive for the performance measurement of public schools. Classification into high, medium, and low support levels could also help to illustrate the relationship between variables and classification levels. More importantly, it could highlight the importance of going beyond single-variable, single-threshold early warning systems (e.g., systems that focus on only one KPI), which overlook complex interactions among predictors. One variable is not sufficient to predict measurements of public school performance. The proposed model based on ANN produces more accurate prediction values than the other existing approaches because of its heuristic learning and correction

technique. The proposed work was developed on the basis of a bio inspirational approach for increasing the performance of the prediction process. ANN assigns weights based on trial and error during the training phase. This proposed work utilised the knowledge of the genetic algorithm to assign the weights of the hidden nodes, and thus its expected outcome and the actual outcome were closely matched. Hence, the proposed model's error rate is very low compared to other algorithms, while its prediction accuracy is also greatly improved.

On the map of the world, Pakistan is facing severe social, demographic, and educational disparities. It is ranked 143rd out of 144 countries on the Global Gender Gap (GGG) index with a score of 0.546, the worst in South Asia [55]. Among South Asian counties, Pakistan's performance in education is not reasonably satisfactory. Moreover, its educational disparities are higher, and significant efforts towards alleviating them have not been observed. Pakistan consists of five provinces, of which Punjab is the most populous. Punjab accounts for more than 56 percent of Pakistan's total population and 52 percent of its gross domestic product. Punjab consists of nine divisions and 36 districts. In Punjab, demographic disparities exist among the various districts [56]. Lahore (its developed district) ranks first and Muzaffargarh (underdeveloped district) last on the Human Development Index. In terms of educational disparities measured in average years of schooling, Muzaffargarh is more deprived, with 4.41 years for males and 1.95 for females, contrary to Lahore, with an average of 8.5 years of education for males and 7.34 years for females.

The same trend is found in all other provinces [57]. One of the probable reasons might be the strong family system in Pakistan, which places all economic responsibility on males, whereas females are not supposed to earn or spend within the family. Hence, education, whose primary purpose is to help secure jobs and livelihoods, might be male-focused. In addition, cultural values in Pakistan do not support the unrestricted mobility of females. They must be accompanied by male members of their families when travelling. Thus, the preferences for educating females are lower within a family. Such values are stronger in rural areas, where education appears to be considered a luxury for girls. Consequently, many females discontinue their education after exhausting the available resources in their hometowns, leading to educational disparities.

The Annual Status of Education Report: Pakistan (ASER-PAK) 2018 presented the current education status in Pakistan in all aspects. Even if we only consider the report for the most advanced province in Pakistan, Punjab, it cited 11% absenteeism among children and 13% among teachers still in public schools. Only 31% of teachers had graduated from an institution, while 59% had obtained professional qualifications or bachelors degrees in education. Regarding school facilities, 79% of public schools had computer labs, and 83% had a library facility. Furthermore, only 2% of primary schools lacked toilets, while 4% were without drinking water. Other factors such as a lack of grants to schools, insufficient classrooms, fewer playgrounds, etc., are also detailed in the report [58]. Such surveys have been performed in the past with attention to specific institutions or regions and considering a limited set of institutional parameters [7,8]. In this research, a maximal set of influencing institutional parameters were included with a broader scope covering the regional level to measure overall, region-wide institutional performance. The results proved that the efficient provision of resources yields better educational results. It was also observed that the urban areas performed well compared to their rural counterparts due to the maximum availability of facilities and resources. Better school infrastructure and physical facilities increased student attendance, strengthened staff motivation, and improved student academic results.

There is always a link between school users (students, teachers) and school architecture. Past studies have demonstrated that a clean and safe learning environment plays a valuable role in academic achievement. Moreover, overcrowding of classrooms, toilets, laboratories, and dormitories, and dilapidated school structures create an uncomfortable school environment. Unhealthy school environments lower the morale of students, teachers, and parents, leading to higher dropout rates and poorer academic achievement [59–61].

Taking the 2030 agenda into consideration, formulating reliable education measures, measuring education disparities among districts, and investigating factors behind education disparities at the household level will all be imperative to the task of recommending effective policy options and the tackling the targets of the Sustainable Development Goals in earnest. This study will help support governance for performance monitoring, policy formulation, target-setting, evaluations, and reforms aimed at addressing the issues and challenges in education worldwide. Gaps in school participation can be better understood in terms of regional socio-economic, demographic, and geographic disparities. There were a few limitations to our study. Firstly, it only covered data for high schools in one province of Pakistan, and the results for other provinces may differ. Secondly, our model utilised a structured dataset, but the results may vary when unstructured or semi-structured data are utilised.

### **6. Conclusions**

Whenever the governmen<sup>t</sup> introduces educational policies that are based on analyses of performance not of a single school but of schools on a massive scale, region-wide—rather than individual–school performance measurements are a practical approach. The level of education in public institutions varies across all regions of Pakistan. The current disparities in access to education in Pakistan are mostly due to systemic regional differences and the distribution of resources. This study, therefore, sought to fill the gaps and emphasise the importance of region-wide measurements of school performance. A machine learningbased method was developed to generate results. It was revealed that aside from student academic results, other factors substantially impact the performance of any school institution. The present study focused on the importance of these other highly influential factors along with student academic results, e.g., teacher–student ratios, the number of schools per region, school locations in rural or urban areas, and the availability of classrooms, electricity in schools, furniture for students, open-air classes, computer labs, science labs, and school playgrounds. Our finding was that in Pakistan, discrepancies in the performance of educational institutions in different regions of the country are due to inequality in the distribution of resources, differences in essential facilities, the number of schools by region, and the influence of school location on motivation, literacy rates, and awareness levels in the local population. This study will help support governance for performance monitoring, policy formulation, target-setting, evaluations, and reforms to address the issues and challenges for education. Moreover, changing socio-economic factors may lead to different results. This research could be conducted on all schools—primary, middle, high—and even institutions of higher learning or in different regions of the nation. In the future, a few advanced ensemble-based machine learning algorithms such as extreme gradient boosting could be utilised in this domain.

**Author Contributions:** Conceptualization, T.M.A., M.M. and K.S.; methodology, T.M.A.,M.M., K.S., I.A.H.; software, I.A.H., M.U.S. and S.L.; validation, M.U.S. and S.L.; formal analysis, T.M.A., M.M. and K.S.; investigation, M.U.S. and S.L.; resources, M.U.S. and S.L.; data curation, T.M.A., M.M., K.S. and I.A.H.; writing—original draft preparation, T.M.A., M.M. and K.S.; writing—review and editing, T.M.A., M.M. and K.S.; visualization, M.U.S. and S.L.; supervision, M.U.S. and S.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** The data associated with this article can be found in the online version at doi:10.17632/637d4s7vjh.1.

**Conflicts of Interest:** The authors declare no conflict of interest.
