Modelling Student Performance in a Structural Steel Graduate-Based Module: A Comparative Analysis Between K-Nearest Neighbor and Dummy Classifiers

Masengo Ilunga, Omphemetse Zimbili, Phahlani Mampilo, Agarwal Abhishek

The predictive strength of the K-Nearest Neighbor (K-NN) and Dummy machine learning classification algorithms is investigated for students' final score. The dependent variable (label) is defined by a binary class, while the different assessments define the independent variables (features). The latter are the module student assessment marks, and the former covers students' final score. The two algorithms have been applied to the Structural Analysis IV, which is an engineering technology module in the Civil Engineering Advanced Diploma, taught at the University of South Africa. Competency level or graduate attribute characterises such a module. The results showed that the accuracy values of K-Nearest Neighbor (K-NN) and Dummy algorithms were 0.95 and 0.79 respectively. However, the values of recall, precision, f1-score, support, kappa coefficient and Matthews correlation coefficient, showed that the Dummy model predicted very poorly the “fail” instances, as opposed to the “pass” instances. Thus, the K-NN classifier outperformed the Dummy classifier. The two algorithms could be simultaneously recommended as guiding tools for academics in predicting students' final score (as fail or pass). However, K-NN is the only algorithm that could be used for both fail and pass.

Full Text