Date of Award
Doctor of Psychology (PhD)
College of Engineering and Computing
Student attrition is one of the long-standing problems facing higher education institutions despite the extensive research that has been undertaken to address it. To increase students’ success and retention rates, there is a need for early alert systems that facilitate the identification of at-risk students so that remedial measures may be taken in time to reduce the risk. However, incorporating ML predictive models into early warning systems face two main challenges: improving the accuracy of timely predictions and the generalizability of predictive models across on-campus and online courses. The goal of this study was to develop and evaluate predictive models that can be applied to on-campus and online courses to predict at-risk students based on data collected from different stages of a course: start of the course, 4th week, 8th week, and 12th week.
In this research, several supervised machine learning algorithms were trained and evaluated on their performance. This study compared the performance of single classifiers (Logistic Regression, Decision Trees, Naïve Bayes, and Artificial Neural Networks) and ensemble classifiers (using bagging and boosting techniques). Their performance was evaluated in term of sensitivity, specificity, and Area Under Curve (AUC). A total of four experiments were conducted based on data collected from different stages of the course. In the first experiment, the classification algorithms were trained and evaluated based on data collected before the beginning of the semester. In the second experiment, the classification algorithms were trained and evaluated based on week-four data. Similarly, in the third and fourth experiments, the classification algorithms were trained and evaluated based on week-eight and week-12 data.
The results demonstrated that ensemble classifiers were able to achieve the highest classification performance in all experiments. Additionally, the results of the generalizability analysis showed that the predictive models were able to attain a similar performance when used to classify on-campus and online students. Moreover, the Extreme Gradient Boosting (XGBoost) classifier was found to be the best performing classifier suited for the at-risk students’ classification problem and was able to achieve an AUC of ≈ 0.89, a sensitivity of ≈ 0.81, and specificity of ≈ 0.81 using data available at the start of a course. Finally, the XGBoost classifier was able to improve by 1% for each subsequent four weeks dataset reaching an AUC of ≈ 0.92, a sensitivity of ≈ 0.84, and specificity of ≈ 0.84 by week 12. While the additional learning management system's (LMS) data helped in improving the prediction accuracy consistently as the course progresses, the improvement was marginal. Such findings suggest that the predictive models can be used to identify at-risk students even in courses that do not make significant use of LMS.
The results of this research demonstrated the usefulness and effectiveness of ML techniques for early identification of at-risk students. Interestingly, it was found that fairly reliable predictions can be made at the start of the semester, which is significant in that help can be provided to at-risk students even before the course starts. Finally, it is hoped that the results of this study advance the understanding of the appropriateness and effectiveness of ML techniques when used for early identification of at-risk students.
Mansour Hamoud Awaji. 2018. Evaluation of Machine Learning Techniques for Early Identification of At-Risk Students. Doctoral dissertation. Nova Southeastern University. Retrieved from NSUWorks, College of Engineering and Computing. (1059)