Analysing Imbalanced Dataset for Postgraduate Student Dropout Using Predictive Analytics

Higher educational institutions face problems with increasing student dropout rates. Past research proposed multiple approaches, such as data mining, to solve this issue. However, the prediction model's ability to determine university students' dropout rate is still in its infancy. We explored various feature selection strategies (filter, wrapper and embedded) on postgraduate studies data from a public university. The dataset is imbalanced, and most of the class is students who finished their studies. The minority class is for students who have not completed their studies. The study applied an oversampling method named SMOTE on the imbalanced data in training to enable the classification algorithm to learn from a balanced dataset, thus minimising overfitting. Five supervised algorithms, Decision Tree, Random Forest, Naive Bayes, Multi-Layer Perceptron and Logistic Regression, were evaluated for the predictive model. The result shows that Random Forest is the best classification algorithm for the dataset. Random Forest produces the highest accuracy value of 100% when using the filter and embedded method and 99.878% when using the wrapper method when using test data.

I. Introduction

An education system consists of hundreds of higher education institutions comprising public and private universities and colleges. These institutions have their own unique identity, missions, and visions. Despite this uniqueness, they still have similarities, such as providing quality education to students and producing high -quality graduates. The education system functions holistically to ensure that the students complete their program and thus produce high-quality graduates for the job market. Unfortunately, some students failed to complete their program. The education system categorises these students as ‘failed’ or ‘dropouts’.