Journal Archive

Open university has contributed to the expansion of higher education and the improvement of excellence of adult learners. However, contrary to expectations, the rate of learners failing to continue their studies and stopping or dropping out of open universities is higher than that of traditional universities(Breslow et al., 2013; Tyler-Smith, 2006). The dropout rate of open universities is 3-4 times that of traditional universities(Lee, Jeong, & Kim, 2020). Students at open universities have characteristics that vary in age, occupation, and academic background compared to students at traditional universities. Unlike traditional students who concentrate on their studies in college, adult learners often play various social roles such as jobs and housework while performing their studies. In addition, difficulties such as a sense of isolation and lack of interaction in a distance education environment may affect continuing study (Kwon, 2010). In order to prevent academic interruption and support academic continuity of open university students, it is necessary to consider new alternatives through consideration of the learning experience of distance university learners(Marr, 2018).

Models of Tinto(1975), Kember(1989), Eaton and Bean(1995) and various empirical studies based on it have continued to identify variables of academic interruption of online university learners and explore the relationship between each variable to explain the phenomenon of academic interruption comprehensively and structurally. Researchers tried to explore various variables such as individual variables such as learners’ environment, psychological variables such as motivation, and educational institutional variables that affect the curriculum and learning to explore the causes of adult learners’ continuity or suspension of study and suggest alternatives(Jung & Lee, 2017).

With digital education platforms on the rise, open universities in many countries have been facing significant challenges related to student dropouts. Studies by Shou et al.(2024), Elibol and Bozkurt(2023), and Appavoo, Gungea, and Sohoraye(2023) indicated that dropout rates are higher in open and distance learning compared to traditional educational formats. These studies emphasized the complexity of accurately predicting and addressing dropout due to the nuanced and multi-faceted nature of student engagement and retention in online environments.

The lack of a consistent definition of “dropout factors” complicates the efforts to address this issue across different educational settings, as highlighted by Elibol and Bozkurt(2023). This inconsistency hampers the ability to develop universally effective interventions, suggesting the need for a more standardized approach to how dropout is understood within the academic community.

Studies on dropouts of domestic cyber college students generally deal with learner variables as important(Im, 2007; Jeon, 2010; Kwon, 2010). Im(2007), who analyzed the causes of dropout by dividing the causes of dropout into maladjustment of learners and discontinuation of purchase of educational services in consideration of the specificity of cyber colleges, reveals that the causes of dropout were in the order of difficulty in securing learning time, financial difficulties, maladjustment to online education environments, and anxiety about cyber colleges’ awareness. Jeon(2010) also identifies variables that predict potential dropouts who are likely to drop out during the degree process based on variables related to dropout of cyber college learners, and suggests implications for them to successfully complete the course. It is revealed that variables directly related to the learning of the subject (total learning time, curriculum/content, and number of lecture accesses) have a relatively important influence on predicting the classification of dropouts of cyber college learners. Kwon(2010) also investigated the effects of personal variables, educational institution variables, and social variables on the decision to drop out intention of cyber universities, and as a result of examining the hierarchical effects of the three variables, only individual variables had a statistically significant effect. Among the individual variables, the variable that had the greatest influence on the intention to drop out of cyber college students was the learner’s interest variable.

Studies have also been conducted on students’ retention and dropout in the Korea National Open University, a remote university that conducts blended learning that combines online education and face-to-face education. Recently, studies have been conducted to reveal the relationship or structural impact between various variables in relation to the discontinuation of studies of adult learners at distant colleges(Kwon et al., 2020; Lee, Jeong, & Kim, 2020). Lee and Kim(2022) analyzed 88,000 data on K University students’ academic continuity and analyzed factors that affect K University students’ academic continuity among their personal characteristics(gender, age, entrance grade, entrance semester) and academic performance (number of subjects participating in mid-term evaluation, final evaluation, rating, acquisition, etc.). Among the variables of students’ personal characteristics, age and admission grade influenced students’ academic continuity. Among the academic performance variables, almost all variables, such as the number of subjects participating in the interim evaluation, the number of subjects participating in the final evaluation, ratings, and acquisition credits, had a relatively large impact on academic continuity. It is interpreted that the variables of academic performance have a great influence on academic continuity because the sincerity of academic performance is linked to academic continuity.

Recent data-based studies have attempted to identify variables related to academic continuation-interruption and reveal a comprehensive relationship, as well as to predict academic continuation or suspension of online university learners(Joung, 2020). Even if the meaning and background of learners’ learning behavior cannot be fully explartificial intelligencened, the product of a certain learning process was used to predict the behavior of learners to continue or stop their studies in the future.

Recently, research and practice on AI-based online higher education have been widely conducted(Ouyang, Zheng, & Jiao, 2022). Studies on AI-based online higher education have been applied to “prediction of learning status, performance or satisfaction, resource recommendation, automatic assessment, and improvement of learning experience”(Ouyang, Zheng, & Jiao, 2022, p. 7893). About two-thirds of cases of applying artificial intelligence to online higher education focus on predicting students’ academic performance, and predicting students’ dropout risks, academy performances, and sites effects in online courses(Ouyang, Zheng, & Jiao, 2022). Mubarak, Cao, and Zhang(2020) built a model to predict students at risk of dropout based on interaction logs in an online learning environment, with an accuracy of 84%. Aguiar et al.(2014) predicted the persistence of online classes by analyzing the portfolio of engineering students and showed results that predict them more accurately than models based on traditional academic data such as SAT scores, GPA, and demographics.

In an online higher education environment with a relatively high risk of dropout, predicting the continuation or suspension of such learning can benefit both individual students and higher education institutions. Baneres, Rodriguez-Gonzalez, and Serra(2019) provided guidance and feedback by early identification of students with the possibility of academic interruption. In this way, the prediction system allows instructors and school administrators to identify problems with students’ studies, support students’ academic processes, and help students continue without stopping their studies(Moreno-Marcos et al., 2018). AI-based predictive modeling, as explored by Shou et al.(2024), provided a potent tool for identifying at-risk students before they leave their courses. By leveraging multidimensional time-series data that include demographic information, learning behaviors, and assessment outcomes, institutions can proactively implement support systems tailored to the needs of individual students. Elibol and Bozkurt(2023) discussed the application of data mining and machine learning techniques in analyzing dropout trends and patterns. These technologies are pivotal in unearthing underlying factors contributing to dropout, which can be obscured in traditional analyses. However, the success of these methods depends heavily on their integration with human-centered educational strategies that address the psychological and motivational needs of students.

Appavoo, Gungea, and Sohoraye(2023) argued for the necessity of induction sessions and enhanced tutorial support to improve student retention. They emphasize that personalized support and tailored intervention strategies are crucial in mitigating the risk of dropout, particularly for students who might feel isolated within the digital learning environment. The combination of AI-driven analytics with human-centered teaching practices offers a comprehensive approach to improving student retention. This integration helps in creating adaptive learning environments that not only predict potential dropouts but also engage students in a manner that is responsive to their individual learning preferences and challenges. The insights provided by these studies underscore the importance of continued research into both technological and pedagogical interventions in open university settings. Future research should focus on refining artificial intelligence models to better predict student behavior and developing frameworks that enhance the human elements of online learning, thereby fostering an educational atmosphere conducive to higher retention rates.

According to Ouyang, Zheng, and Jiao(2022), the most frequently used algorithms in research on AI-based online higher education are decision tree(DT), neural network(NNN), nayve bays(NB), and support vector machine(SVM), and several algorithms are used in one study. When these artificial intelligence-enabled algorithms are applied to multiple variables, the accuracy of the prediction model increases.

AI-based research in online higher education has been increasing rapidly, but the cases applied to open universities are still insignificant. In order to utilize the development of artificial intelligence technology to reduce K University students’ academic dropout and increase academic retention, an artificial intelligence-based K University academic retention prediction system was developed.

Ultimately, this study aimed to derive an analysis model that predicts retention-dropout of K University students by analyzing big data based on artificial intelligence and to provide it to professors and school administrators. In order to develop such a students’ retention analysis model, it was necessary to systematically organize the data and specify the goal of what results they wanted to obtain. The tasks performed before the predictive analysis model can be summarized as follows.

First, the current data status of K University was investigated to provide additional data and consolidate overlapping data. Through this, data indicators were extracted through the expert meeting. Second, we investigated the data status of each administrative department and analyzed their needs to weight the dropout prediction indicators. Third, we conducted FGI with instructors to identify factors of students’ retention.

In order to select indicators for K University’s data status and academic persistence, we investigated all the current data status of K University. Currently, the data status provided by the statistical information system is divided into the following areas: Academic Affairs / Educational Environment / Administration/Major Disclosures / Student Life Cycle Analysis. It supports time series analysis by integrating distributed data within the K University and maintains a system for managing key data of the K University. We also want to provide an easily accessible information environment so that users can utilize the statistics they want in the web environment. We organized the statistical information and confirmed the factors for academic persistence through expert meetings to form the basic variables of the artificial intelligence-based dropout prediction system.

This research investigated the data status of each department and conducted a needs analysis. In fact, the person in charge of each department conducted interviews with each department to investigate all data on campus to secure meaningful data and develop indicators. This research received various opinions and ideas on the areas and forms that need to be analyzed, and sought to explore ways to utilize the analysis results for ‘teaching and learning support services’ and ‘induction of academic continuation’.

The faculty FGI was conducted on teaching and learning aspects, department/university policy-decision aspects, U-Campus data analysis, and survey questions. To select the instructors, we selected professors who are in charge of major policy decisions at the school, professors who have experience as the head of the headquarters of educational informatization, professors who have conducted research on mid- to long-term development strategies, and professors who have served as the director of the Distance Education of Institute. The above professors were provided with explanatory materials on the need and purpose of developing an artificial intelligence-based dropout prediction system, as well as the contents of the data hub portal and learning analytics, and their opinions were collected through in-depth interviews.

After this data restructuring, specific steps were taken for retention-dropout prediction. Data on students’ characteristics(gender, age, region, etc.), affiliation(department, grade, college, etc.), course(taking courses, enrollment, etc.), academic affairs(registration, grade, etc.), learning(online learning variables), and counseling support(counseling, tutoring, etc.) were collected. By synthesizing the data collected from various sources, unnecessary variables were removed and derivatives such as academic participation were generated. After applying the variable selection method, a total of more than 220 variables were finally put in, and these variables are divided into three areas: academic record, registration for courses, and online learning record.

Academic record variables included acquisition credits, the number of F-credit subjects, final evaluation scores, interim evaluation scores, formative evaluation scores, and ratings. Variables of the registration for courses were, for examples, the year of admission, the number of semesters elapsed, the number of leave of absence, the number of unregistered courses, and the number of courses registered. The online learning record variable included learning progress rate, average weekly learning progress, online learning hours, and homepage login count.

A total of 540 million cases were collected, and the data were split up according to the year so that artificial intelligence could train, validate, and test. Among the data from 2018 to 2022, 2020 data, which has different trends in students’ learning from other years due to the impact of the COVID-19 pandemic, was excluded from the analysis. artificial intelligence learned data from 2021 to 2022 and verified it with data in 2019, and tested it with data in 2018. The prediction cycle for retention-dropout of K University students was set on a monthly basis and was designed to predict the likelihood of retention-dropout of students at the beginning of each month based on data up to the previous month.

The importance of variables was analyzed through basic statistical analysis and data preprocessing(abnormalization of outliers and missing values). A variable selection method that selects a variable through an algorithm was applied. Three algorithms (XGBoost, LightGBM, and CatBoost) were used, and the algorithm ensemble technique was applied using the average value of three algorithms.

XGBoost	LightGBM	CatBoost
- Algorithm with parameters added to prevent overfitting - Learning to minimize residuals to reduce the difference between actual and predicted values	- Perform residual minimization learning with vertical asymmetric tree model - Leaf-centric segmentation minimizes predictive error loss values	- Direct use of categorical variables without proactive data preprocessing - Provides categorical loss functions to work effectively for classification problems

The more than 220 variables used in the analysis are classified into three areas (academic record, registration for courses, and online learning record). <Table 3> shows the variables with high importance in predicting the retention-dropout of K University students in each area.

Among the input variables, the most important variable in predicting students’ academic retention was ‘the number of leave of absence’ (importance .613). The second most important variable was ‘total access count in week 6’ (importance .592). In addition, the importance of ‘number of mid-valuation’, ‘number of mid & final evaluation’, and ‘total number of registrations’ were also high.

<Table 3> presents the top 10 variables of high importance according to the three areas of maintenance-dropping prediction modeling of K University students: academic records, course registration, and online learning records. In the prediction modeling of retention-dropout of K University students, ‘number of leave of absence’ in school academic records, ‘number of md-evaluation’ in courses registration, and ‘total number of access in week 6’ in online learning records were the highest. When variables in all areas were input at the same time for prediction modelling, the importance of ‘grade’, ‘total number of registrations’, ‘number of mid & final evaluation’, ‘number of web classes’, and ‘mid-evaluation score’ was high. The importance of variables changed according to the prediction timing and target.

Using data built in consideration of K University students’ admission to graduation, a model was developed to predict students’ retention-dropout by analyzing students’ academic patterns with artificial intelligence. The number of data, accuracy, and recall rate of the prediction model for each algorithm and classification target are presented in <Table 4>. The average accuracy of the prediction model was very good at 97.3% as of June 2023. Among the models, the prediction model for enrolled students using the LGBM algorithm showed the highest accuracy(99.67). Among the various modeling, the modeling for new students using the XGB algorithm showed the lowest accuracy(92.09). Although there are differences between models, the accuracy and recall rate were very good in all models.

The characteristics of dropout between enrolled and new students were different. Students tended to give up their studies when learning activities were low, such as a low progress rate of formative evaluation without submitting assignments. New students were relatively often dropped out even when their learning activities were excellent, such as when submitting assignments and having a high progress rate of formative evaluation.

The developed retention-dropout prediction model for K University students was actually applied to students attending in the spring of 2023, confirming how consistent it is with the actual retention-dropout results and prediction results of students. Based on last month’s data, we operated the prediction model at the beginning of each month, and accordingly, we were able to see how well the predictions of enrolled and expelled students were correct each month.

As of the spring semester of 2023, the actual student prediction rate of enrolled students was 70.33%, and the expulsion prediction rate of expelled students was 89.50%. In addition to the accuracy of the prediction model itself, the retention-dropout prediction for actual K University students in the prediction model as of the end of the spring semester in 2023 showed an accuracy of 79.92% (see <Table 5>). Although the accuracy was very high at the beginning of the semester and the prediction accuracy decreased toward the end of the semester, the overall prediction accuracy was good.

In the retention-dropout prediction, students were classified into high-risk groups (over 70), risk groups (50-69), and general groups (less than 50) according to the possibility of dropout. It was as important to identify students at risk of dropping out as accurately predicting dropout, so high-risk groups and risk groups were set up by applying the actual dropout rate of K University. In 2018-2202, the size of K University students expulsion reached 15-17% every year. Based on information disclosure in 2023, K University’s dropout rate was 20.76%. These statistics of expulsion and subsequent dropout were reflected in the setting of risk and high-risk groups. The student distribution that tracked the predictions of the model considering this average expulsion scale is presented in <Table 6>.

As well as K University students’ retention-dropout prediction models, the interactive dashboards have been developed to provide real-time data on student performance and risk levels. These dashboards are designed to be user-friendly for administrators and faculty, enabling quick identification of at-risk students and facilitating immediate intervention. This tool helps in managing student data more efficiently and allows for swift action to support students in need.

The main page provids the overall monthly retention-dropout prediction results, and the page for each subject provided the retention-dropout prediction results of students taking the course. The prediction status is presented by operating a system that predicts retention-dropout on a monthly basis to instructors, and major influencing factors are identified. It is provided to inquire student prediction information by department and subject.

Risk Group Visualization is another critical feature of our system. Each category (general, at-risk, and high-risk) is clearly displayed with corresponding predictive scores and key indicators that influence their risk status. This visual method helps users quickly understand the risk distribution and focus resources more effectively. Trend analysis over time is conducted through graphical representations of data trends over time. This analysis assists in understanding the effectiveness of interventions and monitoring changes in student behavior or performance throughout the academic cycle. By observing these trends, K University can adjust strategies to better meet the needs of their students. By providing sophisticated visualization tools, K University helps officials make data-based decisions directly that improve student performance and minimize dropout rates. The ability to visualize complex datasets simplifies the decision-making process and supports active educational practices.

Retention-Dropout Prediction : Main	Prediction by Courses

Variables mainly related to students’ retention-dropout were selected, and information on these variables was visualized and presented. It shows how the main variables of the retention-dropout of students vary depending on the student group(new first grade, transferred second grade, transferred third grade). The differences between the levels of major variables and groups can be identified, and necessary measures can be taken accordingly. The The major influencing factors of K University students on their retention-dropout prediction cover ‘number of post-admission semesters’, ‘percentage of total leave of absence’, ‘total unregistered percentage’, ‘number of courses registered(current semester)’, ‘number of subjects retaken’, ‘years of admission’. The values of these factors were presented to the admission grade group and the prediction group(general, risk, high-risk), and the average value of enrolled students was presented so that the levels could be compared together (see <Table 8>).

Number of post-admission semesters	Percentage of total leave of absence
Total unregistered percentage	Number of courses registered - current semester
Number of subjects retaken	Years of admission