A Study on Exploiting Temporal Patterns in Semester Records for Efficient Student Dropout Prediction

Academic achievement data are essential in building a model to predict student dropout. When an attribute of the data has multiple values, each representing a student’s achievement earned over a semester, existing methods typically calculate a mean from those values and use it to build learning data. Such a summary-based approach has been widely used because it can simplify learning processes, including feature extraction. However, model performance can be further improved if patterns in multiple semester values can be properly extracted and used for learning, instead of using summaries. Despite its potential, this problem has not been investigated in previous studies. In this paper, we demonstrate that recurrent neural networks (RNNs) can effectively be used to exploit the patterns in students’ academic records stored by semester. To identify patterns in the data and find solutions suitable for it, various neural network algorithms were compared. Attention was also adopted to improve model performance. Experiments conducted on real student records showed that the gate recurrent unit (GRU) model with multi-head attention achieved an F1 score of 0.9416, which was approximately 5% higher than the existing summary-based approaches. This demonstrates that the semester records exhibit temporal patterns and RNNs can effectively be used to exploit these patterns.