Abstract:

Student’s academic success is very important for the Higher Education Institutions (HEI), impacting the institutions in several ways, e.g., financial, reputation, science, funding, etc. It is a benchmark of the HEI performance. To maximize the students’ ability to succeed, it is common to have in place tutoring actions to support the students with their academic activities, during the academic year, providing personalized advisory and keeping the students engaged with the academic activities.

The EDU.IA project was developed by the University of Trás-os-Montes e Alto Douro (UTAD) with the main goal to enhance the tutoring activities, by using data analytics and Artificial Intelligence (AI) to predict possible difficulties for the students and act as soon as possible. The project’s main focus is on dropout prediction, as this is the most harmful form of academic failure. The knowledge of the dropout probability for each student enables the tutoring program to act preventability and work with the most endangered students. It is thus possible to plan activities and intervene before the dropout becomes a fact.

The EDU.IA platform comprises three elements, in three distinct layers, i.e., data, inference algorithms, and user applications. The data is extracted from the academic records of the students that attended UTAD in the last fifteen years and is integrated in a datawarehouse, designed to provide data to the inference elements. The inference algorithm is an algorithm from the KNearest family that computes the prediction of the future academic performance of a student by classifying the student, according to the comparation of the student’s current grades and those of the students from previous years. The user apps are a web application for tutoring management, which displays a simple list of students and the dropout prediction for each student, for the current academic year. There is also a webservice, mainly for systems integration, so the individual prediction can be integrated in other student related software systems.

A long and demanding work was made to test and choose the most accurate algorithm. It was designed and executed a set of tests, using Artificial Neural Networks, Decision Trees, and Ensemble Methods (Random Forests and Gradient boosting). These algorithms ware able to achieve good results in predicting academic dropout, with emphasis on RF and XGBoost, which demonstrated an accuracy of 88% and 90% in the final test set, respectively.

Although with good accuracy results, the Ensemble Methods require very particular and demanding data preparation, as the university’s digital records are organized in a way that do not favour the easy consumption of data by this type of approach. To overcome this problem, an alternative approach was designed to use a simpler algorithm that could deal with the data constrains and still perform good. The best algorithm for the data available at UTAD is the KNearest algorithm that relies solely in the students’ grades in each curricular unit (course). With this algorithm, it was achieved a dropout prediction accuracy of 80%, which is satisfactory for tutoring management, and is a good trade-off regarding complexity.

In future work, stating this year, we will collect the individual opinion of the tutors regarding their students and compare that data with the inference. We want to know how good is the inference system, when compared with human intuition.

Keywords:

Student dropout, artificial intelligence, dropout prediction, student performance.