STUDENT PREDICTION SYSTEM WITH POTENTIAL DROPOUT USING DATA MINING

• Main Author: Aditya Darmawan, Firdausi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67901
• Institution: Institut Teknologi Bandung

Dropout is one o f the challenges [aced by universities in various world, including Indonesia. Dropout can be reference for how the quality of the academic education system is applied to universities. According to the Higher Education Statistics 2020, there are 600 thousand dropout students in Indonesia, or around 7% of the total 8.5 million. Therefore, universities must have awareness to overcome this dropout risk. In overcoming dropout problems, namely by developing an early warning system for students who have potential to dropout, so that they can intervene ta prevent them before they occur. Early warning is built using data of individual students. One technique that can be xsed is €diicafiona/ Dafa Mining (EDM), which is a technique for evaluating and predicting student performance. EDM is a process of extracting useful information and patterns that are used to predict student performance. In this study, we will ti Io develop prediction system for students who have potential to dropout using data mining algorithms, such as Naive Bayes, S k"M, and ANN. However, there were problems encountered in other previous studies such as poor data quality. Lots of useless or irrelevant data leading to poor performance results. Then modifications will be made Nsfng/eaiore importance and transfer learning to select relevant data and are expected to improve performance results. From the evaltlation results of the three alternative data mining algorithms used, it shows that the ANN has the best precision and rec'all, namely 0.905 and 0.959. Meanwhile, the SVM algorithm only reached 0.892 and 0. 92f) while the Nave Bayes algorithm reac'hed 0.897 and 0.852. The use of ANN wiih modification of feature importance and transfer learning produces an average precision and recall of about 0.921 and 0.967. With these modifications there is an increase in the performance of the model by about 2%.