Predicting Master’s students’ academic performance: an empirical study in Germany
Smart Learning Environments volume 9, Article number: 38 (2022)
The tremendous growth in electronic educational data creates the need to have meaningful information extracted from it. Educational Data Mining (EDM) is an exciting research area that can reveal valuable knowledge from educational databases. This knowledge can be used for many purposes, including identifying dropouts or weak students who need special attention and discovering extraordinary students who can be offered lifetime opportunities. Although former studies in EDM used an extensive range of features for predicting students’ academic achievement (in terms of (i) achieved grades or (ii) passing and failing), those features are sometimes not obtainable for practical usage, and therefore, the prediction models are not feasible for employment. This study uses data mining (DM) algorithms to predict the academic performance of master’ s students by using a non-extensive data set and including only the features that are easy to collect at the beginning of a studying program. To perform this study, we have collected over 700 students' records from 2010 to 2018 from the Faculty of Business Informatics and Mathematics at the University of Mannheim in Germany. Those records include demographics and post-enrollment features such as semester grades. The empirical results show the following: (i) the most significant features for predicting students' academic achievements are the students’ grades in each semester (importance rate between 14 and 36%), followed by the distance from students’ accommodation to university (importance rate between 6 and 18%) and culture (importance rate between 7 and 17%). On the other hand, gender, age, the numbers of failed courses, and the number of registered and unregistered exams per semester are less significant for the predictions. (ii) As expected, predictions performed after the second semester is more accurate than those performed after the first semester. (iii) Unsurprisingly, models that predict two classes yield better results than those that predict three. (iv) Random Forest classifier performs the best in all prediction models (0.77–0.94 accuracy), and using oversampling methods to deal with imbalanced data can significantly improve the performance of DM methods. For future work, we recommend testing the predictive models on other master programs and a larger datasets. Furthermore, we recommend investigating other oversampling approaches.
As the world shifts towards a global economy, individuals and nations have realized that higher levels of education are essential for competitiveness and success. Therefore, pursuing a master’s degree is considered a well-established postgraduate qualification in higher education. It supports building students’ current abilities and help them acquire new skills related to a particular profession. In light of the increasing interest in master's degrees worldwide, failure or drop-out rates are also high. Observation of dropout can take place either from the position of the university institution, which loses a student, or from the viewpoint of the student, who abandons the pursuit of his/her degree. A university dropout is considered a form of academic failure and the necessity of eliminating it is justified by at least four reasons, (i) economic, (ii) social, (iii) individual, and (iv) pedagogical (Staiculescu, 2018). Therefore, many countries have programs in place that promote opportunities to increase the number of highly qualified people for the knowledge society and economy (Kehm et al., 2019). For instance, several projects in Germany have targeted reducing the number of student dropouts as a strategy to enhance the number of professionals who can join the labor market (Mouton et al., 2020). However, the drop-out rate for master’s programs reached 15% for German students and 28% for international students (Kercher, 2018).
Although retention rates of master's students have been widely documented, there are no solid models for predicting students’ success (Rotem et al., 2020). To minimize the wasting of financial and human resources caused by failure or dropouts, it is vital to build models that can predict atrition at the earliest stage possible. Implementing DM methods to educational data is called Educational Data Mining (EDM) (Baker & Yacef, 2009). EDM is a recent research field gaining popularity because of its high potential for improving educational institutions (Baradwaj & Pal, 2011). It concerns developing methods that discover knowledge from educational environment data (Han et al., 2011). It is built from various fields, including data mining (DM) and machine learning, information visualization, computational modeling, psychometrics, and other areas of statistics (Romero & Ventura, 2007). It also concerns social science as it deals with students’ behavior from social and cultural aspects. EDM methods can provide educators and students with valuable insights into the education process, resulting in suitable actions and decisions that improve academic success (Kotsiantis, 2009). The power of EDM can bring numerous advantages. It can help attract, retain, and motivate students’ success. Moreover, it can assist instructors in tracking students’ progress to improve their teaching methods. It can also help students in the process of course selection and educational management. It can also provide students with valuable feedback, offer recommendations, support personalized learning, allocate scholarships, and discover potential Ph.D. candidates.
There are five main methods of EDM (Baker et al., 2011). Those methods are: (i) relationship mining, (ii) prediction, (iii) clustering (iv) distillation of data for human judgment and (v) discovery with models. In our study, we focus on the first two types. To be more precise, there are three types of predictions in higher education: (i) predicting students' academic performance or GPA at a degree level, (ii) predicting students' failure or drop out of a degree, and (iii) predicting students' results in particular courses (Alturki et al., 2020). In this study, we perform the first and second types. Our primary research questions are:
R1 Is it possible to accurately predict the final academic achievement of master’s students?
R1 What attributes have the largest effect on the prediction of students’ academic achievement?
After explaining EDM and introducing our research questions, the rest of the paper is organized as follows: the next section presents related work on predictions performed in higher education. Following that, our research methodology is explained. Afterward, we provide details of the experimental results and discussion. Then, we outline the limitations of this study. Finally, we conclude with a summary of the study's primary outcomes and outline future lines of research.
Literature review on predicting students’ academic performance
According to Rotem et al. (2020), the conducted research regarding students' dropout and postponement at the undergraduate level is more than at the postgraduate level, and no solid predictive models are to be found for postgraduates. For instance, Alemu Yehuala (2015), Aulck et al. (2017), Daud et al. (2017), Pradeep and Thomas, (2015) and Shakeel and Butt (2015) predicted bachelor's degree drop out, Alturki and Alturki (2021), Pal and Pal (2013), Sembiring et al. (2011), Yadav et al. (2011) and Yadav and Pal (2012) predicted bachelor's students' academic achievement at a degree level, and Badr et al. (2016), Huang and Fang (2013), Kovačić (2010) and Osmanbegović et al. (2012) predicted bachelor's students' academic achievement at a course level. The above-mentioned researchers mostly used decision tree algorithms to perform their predictions. They used different type of features. However, gender, age, GPA, income, employment status, and attendance are the most used features.
Based on Nadeem et al., (2021), postgraduate students also face challenges leading to dropout or delay in the program that has remained unexplored. Table 1 summarizes some of the few academic prediction studies that have been performed on a master’s degree level. It compares the different prediction types, the used features in each study, the used algorithms, and the achieved results.
From the studies reviewed in Table 1, most academic prediction studies have been performed on a degree level. On the other hand, only one study by Abu Zohair (2019) was performed on a course level. Moreover, we can notice that predicting the time to complete or not complete a degree is still not common, especially on a master’s level, as we reviewed only one study by Zewotir et al. (2015). When looking at the type of input data that researchers use, we notice that there is a variation from personal and family related features to income and financial aids features. However, post-enrollment features, such as achieved grades are the most common. It can also be seen that some researchers used attributes that are difficult to acquire, such as personality related features by Jeno et al. (2018) and attendance by Yadav et al. (2011). Despite this, other, easily obtainable attributes that could have been relevant have not been included in the previous studies. For instance, culture's influence on academic predictions was not examined in any of the viewed studies. When comparing students' behavior based on their culture, it is essential to know that cultures are typically divided into collectivist and individualist (Moore et al., 2018). Individualist cultures (e.g., people from the USA, Australia, and Europe) impress personal achievement regardless of the expense of group goals, resulting in a strong sense of competition, while collectivist cultures (e.g., people from Pakistan, India, and the Middle East) impress family and team goals over individual requirements (Kim, 1995). Such differences can have a significant impact on students' overall performance. Therefore, it is vital to investigate the impact of culture on academic predictions. Based on the Deutsche Akademische Austauschdienst (DAAD) (Kercher, 2018), the number of international students at German higher education institutions has increased significantly over the past few years, especially in master's courses. This creates the need to investigate the impact of culture in our study. Moreover, none of the studies performed at a master's degree level used the distance from the students' accommodations to the university as a feature for performing academic predictions. Distance increases the financial and personal costs associated with attending classes, which restricts individual choices and leads to low participation rates (Vieira et al., 2018). Consequently, students who live far from campus have a higher likelihood of failing or dropping out. Therefore, in our study, we choose to include distance as one of the predictive features as we believe that it can have an impact on the academic predictions.
This section of the paper presents an overview of the performed study, the type of collected data, the data analysis, the used DM algorithms, and evaluation methods.
The data set of over 700 students used in this study has been obtained from the Business Informatics and Mathematics faculty at the University of Mannheim from 2010 till 2018. It should be noted that we have followed the European data protection regulations for performing this study and all the collected records have been anonymized prior to working on them. For the purpose of ensuring the reliability of the data, we excluded those students who had not completed their degrees prior to the Covid-19 pandemic. Among the reasons for the exclusion are the drastic changes in the nature of examinations and learning styles, e.g., online exams.
The Business Informatics master’s program's intended duration is four semesters (two years) with approx. 120 European Credit Transfer System (ECTS). However, it usually takes up to six semesters. Students in German universities have the option to register or unregister for course examinations each semester. Therefore, it is common that students postpone an examination for the next semester or the one after. The number of registered exams represents the amount of studying load, i.e., the more registered exams, the more the load is on the student. For our prediction study, we intend to select only easy-to-collect attributes that can be collected from any university database, as shown in Table 2. We have used a combination of demographics and post-enrollment features.
Before performing the academic achievement predictions, it is essential to analyze the dataset at hand. As shown in Fig. 1, the number of male students significantly exceeds the number of females. Furthermore, the number of students coming from individualist cultures slightly exceeds those from collectivistic cultures. We can also notice that most enrolled students are 24 and 23 years old, and only very few are in their thirties. Regarding students’ performance (Figs. 2 and 3), we can see that most enrolled students passed the master’s program. However, a considerable amount of failure and dropout needs to be given attention. Moreover, the “Above average” students represent the largest number of students, followed by the “Average” students, then finally the “Below average” students.
This study has been carried out on the Anaconda 4.13.0 (a free OS-independent platform) distribution with Python version 3.8.8. Amongst the Python libraries used in this study are Scikit-learn (for ML algorithms), Pandas (to import and build Data Frames), NumPy (for array computing), and imblearn (for imbalanced data manipulation).
Several DM algorithms can be used to predict the students' graduation performance or dropout. However, the literature review suggests that, in general, there is no single DM method that works best in all contexts. Following are the six DM methods that have been applied in this study:
Logistic Regression (LR) A supervised DM algorithm that attempts to distinguish between classes (or categories) by analyzing the relationship between existing independent features (Geng, 2006). In our study, we use the Binary logistic regression in the cases where the dependent feature has only two possible outcomes and Multinomial logistic regression, where the dependent feature has three possible outcomes.
Random Forest (RF) A supervised DM algorithm that builds multiple decision trees and merges them. It uses voting mechanisms from the multiple decision trees to improve the shortcomings of a single DT and get more accurate predictions (Breiman, 2001). Each tree in the random forest provides a class prediction, and the class with the most votes become the model’s prediction.
K-Nearest Neighbor (KNN) A supervised DM algorithm for estimating the likelihood that a data point will become part of one group based on measuring the distance between the classified instance and the closest training examples in the feature space (Clark, 2013).
Naïve Bayes (NB) A supervised DM algorithm that assumes that the features are independent of each other (Harrington, 2011). It is based on the Bayes theorem, which states that if event B has happened, then we can find the probability of event A, and represented as follows: P(A|B) = (P(B|A)* P(A))/P(B).
Support Vector Machine (SVM) A supervised DM algorithm that seeks to find the hyperplane best separating the data points in high dimensional space by maximizing the margin (Clark, 2013).
Artificial Neural Networks (ANN) A series of algorithms that endeavor to recognize underlying relationships in a set of data by mimicking the information process of the human brain (Clark, 2013). It takes place in two phases. First, the network is trained on paired data to determine the input–output mapping. Then, the weights of the connections between neurons are fixed, and the network is used to determine the classifications of a new set of data.
Cross-validation is mainly used in settings where the goal is prediction, and one wants to estimate how accurately a predictive model will perform in practice. Our study evaluates the prediction models using non-exhaustive cross-validation (k-fold cross-validation). In each cross-validation method, we use four evaluation measures which are precision, recall, F1 score, and overall accuracy, explained as follows:
Precision: the ratio of correctly predicted positive observations to the overall predicted positive observations. It is calculated as precision = (TP)/(TP + FP).
Recall: the ratio of correctly predicted positive observations to the total observations in an actual class. It is calculated as recall = (TP)/(TP + FN).
F1 score: the weighted average of Precision and Recall. It is calculated as F1 score = (2 * Precision * Recall)/(Precision + Recall).
Accuracy: the correctness of value, i.e., the ratio of correctly predicted observation to the total observations. It is calculated as accuracy = (TP + TN)/(TP + TN + FP + FN).
where: TP = True positive; FP = False positive; TN = True negative; FN = False-negative.
In this section of the paper, we present the results obtained from using the six DM algorithms that have been previously described in “DM algorithms” section. For evaluating the performance, and as we have a small dataset, we used tenfold cross-validation (each time, nine of the folds are used for training and one fold is used for testing the model, and the holdout method is repeated ten times).
Predicting students’ academic achievement
This section provides an overview of the results achieved from using traditional DM algorithms to perform academic predictions. For the predictions performed after the first studying semester, we have selected (1) Gender, (2) Enrollment_age, (3) Culture, (4) Distance, (5) Grade_sem1, (6) F_sem1, (7) Unregistered_exams1, and (8) Registered_exams1 as predictors. For performing the predictions after the second demester, we have selected, (1) Gender, (2) Enrollment_age, (3) Culture, (4) Distance, (5) grade_sem1, (6) grade_sem2, (7) F_semester1, (8) F_semester2, (9) Unregistered_exams1, (10) Unregistered_exams2, (11) Registered_exams1, and (12) Registered_exams2 as predictors.
Predict students' academic_status ("Completed" or "Not_completed"):
Table 3 compares the performances of the different DM algorithms that have been used for predicting students’ academic status. We can notice that all the DM algorithms generally provided good accuracy. However, and unsurprisingly, the accuracy is always better when performing the prediction after the second semester. We can also notice that the models best predict the “completed” students (which represents the majority class). For instance, in the case of LR, the precision, recall, and the F1 score reached 0.91, 0.98, and 0.95, respectively, for the “Completed” class. On the other hand, the precision, recall, and F1 score are 0.48, 0.14, and 0.21 for the “Not_completed” class.
Predict students’ academic_grade (“Above average”, “Average”, or “Below average”)
Table 4 compares the performances of the different DM algorithms used to predict the academic grade after the first and second studying semesters. Just like the previous cases, the model works best in predicting the majority class, which is the “Above_average” students, in this case, followed by the second major class (“Average”).
Dealing with imbalanced datasets using SMOTE
By viewing the results in “Predicting Students’ academic achievement” section, one can notice that all the classifiers achieved high accuracy. However, low precession, recall, and F1 score for the minority classes. These misleading results are typical when analyzing imbalanced data. Several techniques have been proposed to solve the problems associated with learning from imbalanced data. Those techniques are (i) resampling (by either oversampling the minority class or under-sampling the majority class), (ii) feature selection, and (iii) cost-sensitive learning. Since we have a limited dataset and a small number of features, over-sampling is the optimal approach. Over-sampling simulates data points to enhance balance across the classes. There are several over-sampling techniques. Our study explores using Synthetic Minority Oversampling Technique (SMOTE), which was proposed to improve random oversampling as it overcomes the overfitting problem posed by random oversampling (Chawla et al., 2002). SMOTE synthesizes new minority instances between existing (real) minority instances. These synthetic training records are generated by selecting one or more of the k-nearest neighbors for each example in the minority class. Then, the data is generated by randomly choosing the features between those two data points. After the oversampling process, the data is reconstructed, and the classification models can be applied to the processed data. Tables 5 and 6 below show the significant improvements in predicting the minority classes after applying SMOTE. For instance, the minority class F1 score using LR raised by 60% after the first semester and 47% after the second semester. As for RF, the F1 score for the minority class raised by 44% and 39% after the first and second semester, respectively. Also, KNN raised by 39% and 36% after the first and second semester, respectively.
Feature importance on the overall prediction
Feature Importance refers to the techniques that calculate a score to each input feature for a given model where the scores represent the “importance” of each feature. A higher score means that the specific feature will have a more significant effect on the predictive model. There are various functions for generating feature importance. However, since Random Forest provided the best accuracy, it is reasonable to find the impact of each feature on the predictions performed by that classifier. Therefore, we use the Random forest permutation importance measurement, which was introduced by Breiman (2001). The feature selections are performed by looping through each column in the dataset while making predictions, shuffles the column, and making predictions with the shuffled column. If a column is significant to making predictions, shuffling that particular column should increase the error term and vice-versa. Therefore, those columns that lead to a maximum increase in error (loss function) are considered the most important.
Feature importance on the predictions performed after the 1st semester
By viewing Table 7, one can notice that the most significant attribute for performing the predictions after the first semester is “Grade_sem1” followed by “Distance”, then “Culture”. Moreover, “Registered_exams1” and “F_sem1” have a minor impact. On the other hand, “Gender”, “Enrollment_age”, “Unregistered_exams1” have the most negligible impact on the prediction.
Feature importance on the predictions performed after the second semester
By viewing Table 8, one can see that the most significant attributes for performing the predictions after the second semester are “Grade_sem2” followed by “Grade_sem1”. While “Culture” and “Distance” have some effect on the prediction, the rest of the features have no significant impact.
Discussion of the results
This section answers our research questions and discusses the results obtained from the predictive models. Our first research question was whether it is possible to accurately predict the academic achievement of master’s students at an early stage. We have built four initial models to answer this research question; two are designed to make predictions after the first studying semester, and two are designed to perform the predictions after the second semester. By going back to Tables 3 and 4, we can notice that the results of predicting the largest classes (“Complete” and “Above_average”) are better than the rest of the classes (“Not_completed”, “Average”, and “Below_average”). This finding was also reported by Nguyen Thai Nghe et al. (2007) which stated that the accuracy of the majority class is higher in their academic prediction study. The minority class is always more challenging to predict because there are only few examples of this class to train on.
As a general trend, the predictions performed after the second semester yield more significant results than those performed after the first semester (see Fig. 4). In our view, it is reasonable since after the second semester we have a more realistic picture of students' performance than we do at the end of the first semester. We can also notice that predicting the academic status (which comprises two classes) is more accurate than predicting the graduation grade (which comprises three classes). This supports the findings of Nghe et al. (2007), who reported that predicting two-class problems produces more accurate results than predicting three or more class problems (i.e., the more the classes, the more challenging the prediction is). To get into more details regarding the performance of the classifiers, we can see that they gave similar accuracies, with RF performing the best in most cases (0.92- 0.94 in the cases predicting the academic status and 0.77- 0.81 in the cases predicting the graduation grade). This is similar to the results of Zhao et al. (2020) who reported that RF performed the best among seven other algorithms. This is not surprising because RF is an ensemble algorithm that uses bagging as the ensemble method and decision trees as the individual model. Ensemble algorithms can be more accurate than single models as they tend to repeat the process many times such that the model learns the data and makes proper predictions. Another reason behind the excellent performance of RF is that it chooses features randomly during the training process. Therefore, it does not depend highly on any specific set of features. This randomized feature selection is a unique characteristic of RF.
After briefly discussing the accuracy of the initial models, the question that comes to mind is whether the models are reliable for practical usage. Although we achieved high prediction accuracy in all four models, they are misleading results and unreliable for implementation. That is because other evaluation methods such as the precision, recall, and F1 score for the minority classes are not sufficient enough. We worked on that issue by using SMOTE (Tables 5 and 6). Figure 5 below is an example that compares the F1score of the minority class (Not_completed) before using SMOTE and after using it. Although the accuracy of the classifiers slightly decreased, they are more applicable as we were able to have high precision, recall, and F1 score for the minority classes. We can also notice that in all four models, the RF classifier continues to perform the best compared to the rest of the DM algorithms that have been explored in this study (0.82- 0.92 accuracy).
The second research question in this study is regarding finding out which attributes have the most effect on the prediction of students’ academic achievement. As previous research has shown that personal features, such as marital status, have an impact on student performance, many researchers such as Nghe et al. (2007) and Zhao et al. (2020) examined such features on predicting students’ academic achievements. Moreover, personality features such as motivation and competence, are known to have a strong effect on students’ achievement as proved by Jeno et al. (2018). However, such features are challenging to collect, i.e., it is not common for academic institutes to collect such data during the enrollment process. Therefore, we find that using such features is not effective for generalizing the predictive models and testing them in different academic institutions. In this study, we seek to perform academic predictions using only easy to collect attributes that are available at any university database. Using this approach allows us to test our models across different academic disciplines and facilities.
We have found that the most crucial attributes for performing the predictions after both the first and the second semester are the achieved academic grades in each semester (importance rate: 14–36%). This supports the finding of Rotem et al. (2020) and Asif et al. (2017). The second most important features are culture and distance, with an importance of 7–17% and 6–18%, respectively. Even though the distance between students' accommodations and universities has been a subject of interest for researchers and universities for decades (Simpeh & Akinlolu, 2018), it has not yet become common to use it as a predictor for performing academic predictions. Students who live far from campus are more likely to fail or dropout as it is difficult for them to attend classes on a regular basis. In addition, they are less likely to make use of university facilities (e.g., library) for an extended period of study time.
As discussed in the literature, cultures' behavior toward learning may differ. Generally speaking, students from individualistic cultures tend to have a higher desire to compete with themselves and with others. Therefore, they have higher chances of succeeding the master’s program. However, there are other factors that can influence international students (which are “collectivistic” in our study) to dropout from German educational programs. Those factors include poor linguistic proficiency, financial problems, lack of social and academic integration, and misconceptions regarding the teaching and learning culture at German higher education institutions (Kercher, 2018).
Attrition rates are known to be higher among students who failed courses than among those who did not, and the rates increase as the number of failed courses increase (Ajjawi et al., 2020). In our study, the number of failed courses were surprisingly found to have a minor effect on predicting the academic status and the graduation grade (with an importance rate of 4–10%). This contradicts the findings of Alturki and Alturki (2021) and Kabakchieva (2013), who observed that the number of failed courses is essential for predicting bachelor's students' achievement. This inconsistency may be because course failure in master's programs is not as common as in bachelor's programs.
Balancing the academic load is vital to students' academic achievement (Alturki et al., 2020). In fact, Alemu Yehuala (2015) found that it is one of the main significant attributes for predicting academic achievement. We tested this theory by investigating the impact of the number of registered and unregistered exams per semester. We found that the number of registered exams has a minor effect with a 5–10% importance rate. As for the number of unregistered exams, it has even a lower effect (2–6%) compared to the rest of the post-enrollment features.
Moreover, we found that enrollment age has almost no effect (4–9%). This finding is in line with the findings of Kovačić (2010). However, it contradicts the finding of Zewotir et al. (2015) who observed that age matters in predicting master students time to graduate or dropout. In our case, it does not surprise us that age did not influence the prediction as there is no significant gap between the ages of most applicants (most students are in their twenties). Researchers have conducted studies at different levels throughout the world that have shown a significant difference in academic performance between males and females. Therefore, gender has been used the most in the literature compared to other demographics in predicting academic achievement (Alturki et al., 2020). In our case, gender does not affect the prediction, as it had an importance rate of only 2–4%. This is also in line with the findings of Kovačić (2010), Osmanbegović et al. (2012), and Zewotir et al. (2015).
Due to time constraints, this study was only performed on one studying program (Business Informatics). However, using the same prediction models in different master’ s programs could give us more insights into whether the predictive models could be generalized and sufficiently work for other programs.
Ensemble methods are known for avoiding overfitting and improving predictions. In this study, we used Random Forests, which is a Bagging (or bootstrap aggregation) method. However, other ensemble methods, such as boosting and stacking, are worth exploring.
Although there are many approaches to dealing with imbalanced datasets, oversampling techniques are optimal for our predictive models. Therefore, we chose to explore SMOTE. However, other oversampling techniques such as adaptive synthetic sampling (ADASYN) and Data augmentation can also be explored.
Conclusion and future works
Collecting a wide range of student data, other than academic performance, for instance, students' health issues and workload, if employed, can be beneficial for predicting students' academic achievement and reducing students’ dropout. However, the downside is that they are expensive to gather. Therefore, the main objective of this study was to perform academic predictions at a master’ s degree level using only the data that can be easily collected at the beginning of a studying program. We compared the performance of six classifiers, namely: LR, RF, NB, KNN, SVM, and ANN, in predicting students' academic performance and explored using SMOTE to improve our predictions and deal with our imbalanced dataset. Results from our prediction models reveal that it is possible to predict academic achievement with a high accuracy using only a small set of features. Those results can assist in building an early warning system. Such a system will allow instructors to know the students at risk of dropping out or those with higher chances of failure and need to be given special attention. It will help academic institutions increase the academic success and reduce the financial loss caused by students’ dropout or failure. However, in order to benefit the most from such system, it must be carefully constructed and continuously monitored.
Future studies should expand on our study by performing earlier predictions (prior to enrollment) as this can bring more benefits to the educational society. Alturki and Stuckenschmidt, (2021) suggested that earlier predictions can be achieved using pre-enrollment tests. Moreover, we highly endorse testing the predictive modes on other master programs. Also, future studies should explore more ensemble techniques to perform academic predictions. We also suggest investigating other oversampling techniques to deal with the imbalanced datasets. Finally, we highly encourage educators and researchers to apply more EDM studies to postgraduate students to have a more realistic comparison between undergraduate and postgraduate EDM studies.
Availability of data and materials
More data is available from the corresponding author on reasonable request.
Deutsche Akademische Austauschdienst
European credit transfer system
Educational data mining
Artificial neural networks
Synthetic minority oversampling technique
Support vector machine
Abu, L. M. Z. (2019). Prediction of student’s performance by modelling small dataset size. International Journal of Educational Technology in Higher Education, 16(1), 1–18. https://doi.org/10.1186/S41239-019-0160-3/FIGURES/13
Ajjawi, R., Dracup, M., Zacharias, N., Bennett, S., & Boud, D. (2020). Persisting students’ explanations of and emotional responses to academic failure. Higher Education Research and Development, 39(2), 185–199. https://doi.org/10.1080/07294360.2019.1664999
Alemu, Y. M. (2015). Application of data mining techniques for student success and failure prediction (the case of Debre_Markos University). International Journal of Scientific and Technology Research, 4(04), 666.
Alturki, S., & Alturki, N. (2021). Using educational data mining to predict Students’ academic performance for applying early interventions. Journal of Information Technology Education: Innovations in Practice, 20, 121–137.
Alturki, S., Hulpus, I., & Stuckenschmidt, H. (2020). Predicting academic outcomes: A survey from 2007 till 2018. Technology, Knowledge and Learning. https://doi.org/10.1007/s10758-020-09476-0
Alturki, S., & Stuckenschmidt, H. (2021). Assessing students’ self-assessment ability in an interdisciplinary domain. Journal of Applied Research in Higher Education, ahead-of-print. https://doi.org/10.1108/JARHE-01-2021-0034/FULL/XML
Asif, R., Merceron, A., Abbas, A. S., & Ghani, H. N. E. D. N. (2017). Analyzing undergraduate students’ performance using educational data mining. Computers and Education, 113, 177–194. https://doi.org/10.1016/j.compedu.2017.05.007
Aulck, L., Velagapudi, N., Blumenstock, J. & West, J. (2017). Predicting Student Dropout in Higher Education. https://arxiv.org/pdf/1606.06364.pdf.
Badr, G., Algobail, A., Almutairi, H., & Almutery, M. (2016). Predicting students’ performance in University courses: A case study and tool in KSU Mathematics Department. Procedia Computer Science, 82, 80–89. https://doi.org/10.1016/j.procs.2016.04.012
Baker, R., & Yacef, K. (2009). The state of educational data mining in 2009: A review and future visions. Journal of Educational Data Mining, 1(1), 3–17.
Baker, R., Isotani, S., & Carvalho, A. (2011). Mineração de Dados Educacionais: Oportunidades Para o Brasil. Revista Brasileira de Informática Na Educação. https://doi.org/10.5753/rbie.2011.19.02.03
Baradwaj, B. K. & Pal, S. (2011). Mining educational data to analyze students” performance. In IJACSA) international journal of advanced computer science and applications (Vol. 2, Issue 6). www.ijacsa.thesai.org.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324.
Calisir, F., Basak, E., & Comertoglu, S. (2016). Predicting academic performance of Master’s Students in engineering management. College Student Journal, 50(4), 501–512.
Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16, 321–357. https://doi.org/10.1613/JAIR.953.
Clark, M. (2013). An introduction to machine learning with applications in R. Retrieved 17 October, 2022 from http://web.ipac.caltech.edu/staff/fmasci/home/astro_refs/ML_inR.pdf.
Daud, A., Aljohani, N. R., Ayaz, A. R., Lytras, M. D., Abbas, F. & Alowibdi, J. S. (2017). Predicting student performance using advanced learning analytics. In WWW 17 Companion: Proceedings of the 26th International Conference on World Wide Web Companion. https://doi.org/10.1145/3041021.3054164.
Geng, M. 2006. A comparison of logistic regression to random forests for exploring differences in risk factors associated with stage at diagnosis between black and white colon cancer patients. Master's Thesis, University of Pittsburgh. Retrieved 17 October, 2022 from http://d-scholarship.pitt.edu/id/eprint/7034.
Han, J., Kamber, M. & Pei, J. (2011). Data Mining. Concepts and Techniques, 3rd Edition (The Morgan Kaufmann Series in Data Management Systems).
Huang, S., & Fang, N. (2013). Predicting Student Academic Performance in an Engineering Dynamics Course: A Comparison of Four Types of Predictive Mathematical Models. https://doi.org/10.1016/j.compedu.2012.08.015
Jeno, L. M., Danielsen, A. G., & Raaheim, A. (2018). Educational Psychology an International Journal of Experimental Educational Psychology A Prospective Investigation of Students’ Academic Achievement and Dropout in Higher Education: A Self-Determination Theory Approach. https://doi.org/10.1080/01443410.2018.1502412
Kabakchieva, D. (2013). Predicting student performance by using data mining methods for classification. Cybernetics and Information Technologies, 13(1), 61–72. https://doi.org/10.2478/cait-2013-0006
Kehm, B. M., Larsen, M. R., & Sommersel, H. B. (2019). Student dropout from universities in Europe: A review of empirical literature. Hungarian Educational Research Journal, 9(2), 147–164. https://doi.org/10.1556/063.9.2019.1.18
Kercher, J. (2018). Academic success and dropout among international students in Germany and other major host countries.
Kim, U. 1995. Individualism and collectivism a psychological, cultural and ecological analysis. http://eurasia.nias.ku.dk/publications/.
Kotsiantis, S. (2009). Educational data mining: A case study for predicting dropout-prone students. International Journal of Knowledge Engineering and Soft Data Paradigms, 1(2), 101. https://doi.org/10.1504/ijkesdp.2009.022718
Kovačić, Z. J. (2010). Early prediction of student success: Mining students enrolment data. In Proceedings of informing science and IT education conference (pp. 647–665).
Moore, D. A., Dev, A. S., & Goncharova, E. Y. (2018). Overconfidence across cultures. Collabra: Psychology, 4(1), 36. https://doi.org/10.1525/collabra.153.
Mouton, D., Zhang, H., & Ertl, B. (2020). German university student’s reasons for dropout. Identifying latent classes. Journal for Educational Research, 12(2), 190–224.
Nadeem, M., Palaniappan, S., & Haider, W. (2021). Impact of Postgraduate Students dropout and delay in University: Analysis using machine learning algorithms. International Journal of Advanced Trends in Computer Science and Engineering, 10(3), 1821–1826. https://doi.org/10.30534/ijatcse/2021/461032021
Nguyen, T. N., Janecek, P. & Haddawy, P. (2007). A comparative analysis of techniques for predicting academic performance. In 2007 37th annual frontiers in education conference-global engineering: Knowledge without borders, opportunities without passports, T2G-7-T2G-12. https://doi.org/10.1109/FIE.2007.4417993.
Osmanbegović, E., Suljic, M., & Suljić, M. (2012). Data mining approach for predicting student performance. Journal of Economics and Business, 10(1), 958.
Pal, A. K., & Pal, S. (2013). Analysis and mining of educational data for predicting the performance of students. International Journal of Electronics Communication and Computer Engineering, 4(5), 2278–4209.
Pradeep, A., & Thomas, J. (2015). Predicting college students dropout using EDM techniques. International Journal of Computer Applications, 123(5), 26–34.
Romero, C., & Ventura, S. (2007). Educational data mining: A survey from 1995 to 2005. ScienceDirect, 33(33), 134–146. https://doi.org/10.1016/j.eswa.2006.04.005
Rotem, N., Yair, G., & Shustak, E. (2020). Dropping out of Master’s Degrees: Objective Predictors and Subjective Reasons., 40(5), 1070–1084. https://doi.org/10.1080/07294360.2020.1799951
Sembiring, S., Zarlis, M., Hartama, D., Wani, E., & Magister, P. (2011). Prediction of student academic performance by an application of data mining techniques. International Conference on Management and Artificial Intelligence, 6, 110–114.
Shakeel, K. & Anwer, B. N. (2015). Educational data mining to reduce student dropout rate by using classification. In 253rd OMICS international conference on big data analysis & data mining.
Simpeh, F. & Akinlolu, M. (2018). Importance level of on-campus student housing facility spaces: Perception of Postgraduate Students. In 10th Cidb postgraduate conference
Staiculescu, C. & Richiteanu, N. E. R. (2018). University dropout. Causes and solution. In Mental health global challenges XXI Century conference proceedings.
Vieira, C., Vieira, I., & Raposo, L. (2018). Distance and academic performance in higher education. Spatial Economic Analysis, 13(1), 60–79. https://doi.org/10.1080/17421772.2017.1369146
Yadav, S. K., Bharadwaj, B., & Pal, S. (2011). Data mining applications: A comparative study for predicting student’s performance. International Journal of Innovative Technology and Creative Engineering, 1(12), 13–19.
Yadav, S. K., & Pal, S. (2012). Data mining: A prediction for performance improvement of engineering students using classification. World of Computer Science and Information Technology Journal (WCSIT), 2(2), 51–56.
Zhao, Y., Qiangwen, X., Ming, C., & Gary, M. W. (2020). Predicting student performance in a Master of data science program using admissions data. In Proceedings of The 13th international conference on educational data mining (pp. 325–33).
Zimmermann, J., Brodersen, K. H., Pellet, J.-P., August, E. & Buhmann, J. M. (2011). Predicting graduate-level performance from undergraduate achievements. In Proceedings of the 4th international conference on educational data mining. https://www.researchgate.net/publication/221570422_Predicting_Graduate-level_Performance_from_Undergraduate_Achievements.
Zewotir, T., North, D., & Murray, M. (2015). The time to degree or dropout amongst full-time Master’s students at University of KwaZulu-Natal. South African Journal of Science, 111(9/10), 1–6. https://doi.org/10.17159/SAJS.2015/20140298
The authors declare that there is no competing interest.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
Alturki, S., Cohausz, L. & Stuckenschmidt, H. Predicting Master’s students’ academic performance: an empirical study in Germany. Smart Learn. Environ. 9, 38 (2022). https://doi.org/10.1186/s40561-022-00220-y