Abstract:

Predicting students at risk is one of the key problems in higher education. Poor academic performance has a significant impact on the society and the economy in general. For example, college dropouts often accumulate loans that puts them in serious financial burden and can potentially impact their physical and mental health. If one is able to predict a student at risk in advance, then mediating action can be taken to preempt failure and to prevent possible dropout. However, this is a difficult problem because a students’ performance in the next semester depends not only on their current cumulative grade point average (GPA), but also on the particular sequence of courses they have attempted in their performance in these courses. For example, a student’s weak performance in the Physics sequence will clearly have an impact on their performance in a Circuits course that depends heavily on Physics at a later stage. In addition, external factors like high-school scores, race, gender, full-time vs. part time status, and socio-economic factors may have an impact as well. A variety of machine learning algorithms have been used for identifying students at risk in the past including Naïve Bayes, Logistic Regression, Random Forest, Ada Boost and Boosted Decision Trees with reasonable performance in predicting dropouts or higher risk students ranging from an accuracy from 70% to 93%. Some previous approaches considered only the academic performance of students while other considered a wider context. In this paper we apply various deep learning approaches to address the problem of predicting students at risk based primarily on their performance in prior semesters. Data from undergraduate students in a college of engineering from 2014 to 2019 at an American University was used. After cleaning, the data included over 3,800 students taking over 1500 courses taught by 160 instructors over 20 terms. Data from Computer Engineering, Computer Science, Electrical Engineering, Civil Engineering, Mechanical Engineering, Chemical Engineering programs were considered. The data included accumulated hours, GPA in each course taken, instructor for each course, academic status, and the cohort. Only the engineering courses were included. Various approaches were explored. A 1D Convolutional Neural Network (CNN) was used to model and predict the students’ GPA in the next Semester. Recurrent Neural Networks (RNN) using Gated Recurrent Units (GRUs) were also evaluated. Long short-term memory (LSTM) networks have been recently successful in prediction tasks based on time series data in a variety of domains. LSTM network was consequently also used to predict the GPA of a student in the next semester based on history of the courses being taken. The models were evaluated using 10-fold validation. The results from various models are compared using F1-measure, ROC and Precision-Recall curves.

Keywords:

Deep learning, at-risk students, higher education, dropout.