Early Prediction of Student Performance in Blended Learning Courses Using Deep Neural Networks

In this paper, we experimented on developing prediction models for student performance in early stages of blended learning courses using deep neural network (NN) architecture and utilizing online activity attributes as input patterns. The online activity attributes were extracted from the activity logs stored by Moodle. A total of 885 records from undergraduate students taking three 3 different courses under 16 different classes were utilized. First, a series of experiments was conducted to determine the hyperparameters for a top performing NN model which then served as baseline classifier. Afterward, experiments were conducted to test the performance of the model for predicting student outcomes (pass or fail) both for the midterm and finals period using activity data generated prior to the midterm period. Results indicate that only low prediction performance can be achieved at an early stage, more specifically during the first month of the course. However, both accuracy, as well as ROCAUC score, improves as more data is accumulated up to the third month. This result supports the findings from previous studies. The highest accuracy achieved for predicting finals outcomes for a single course is 91. 07% with ROCAUC score of 0. 88 while for midterm outcomes the highest is 80. 36% accuracy with ROCAUC score of 0. 70. This study is a part of an ongoing work that aims to develop a tool that can be applied in selected blended learning dimensions to provide a basis for automatic feedback and instructor support.