Mining for Marks: A Comparison of Classification Algorithms when Predicting Academic Performance to Identify "Students at Risk"

A major concern for higher education institutions is the high failure and drop-out rates amongst students, especially first year students. Tertiary institutions thus have a common interest in identifying students at risk of failing or dropping out. Previous research studies have identified factors that influence success/failure which include, but are not limited to, the students' personal information, academic background and social environment. This study aims to use the emerging field of Educational Data Mining as a preventative measure rather than reiterate factors that influence success. The first year student data collected and stored in the School of Computer Science at the University of the Witwatersrand has been utilised in this study. The study used the students' first semester/midyear mark to predict success/failure at the end of the academic year. This will assist in identifying students at risk of failing and could assist with early intervention. A modified version of the CRISP-DM methodology was used. The investigation was broken down into two phases: training and test phase. In the training phase, student data from the years 2009 to 2011 were modelled using the WEKA Explorer GUI. Three classifiers: J48 classifier, Naïve Bayes and Decision Table, were used for modelling and were also compared. Using both the run information from WEKA and performance metrics, the J48 classifier was shown to be the better performing algorithm in the training phase. This algorithm was then integrated into the back-end of the Success Or Failure Determiner (SOFD) tool, which was created specifically for this study. In the test phase 92% of the instances were predicted correctly. Furthermore 23 of the 25 students who failed were flagged. The research findings indicated that the midyear mark can be considered as a factor which correctly predicts the Computer Science I final year marks. After further investigation with larger sample sizes, the tool can be used practically in the school of Computer Science to identify students at risk of failing.