Comparing classification models in the final exam performance prediction

The use of Learning Management Systems (LMS) and the web-based formative and summative assessments during the traditional teaching in classroom provides the huge amount of data on students´ behavior and results at the point of time when the course is still in progress. This data could be used for the final exam performance prediction so that the excellent as well as the students requiring help could be detected. The data on 302 students enrolled into the first year Physics course of a biomedical university study program in 2011/2012 have been used. The data were preprocessed by dividing Croatian grading system which comprises 5 grades (range 1-5; 1=fail, 5=excellent) into 2 categorical classes in one version of the experiment and into 3 categorical classes in the second version. Several up to date algorithms for classification were applied without and with attributes optimization by a genetic algorithm. Also, since students have 4 chances at 5 exam terms during an academic year, 3 different dependent variables were constructed. In order to evaluate the performance that each of the classification models had the following performance criteria were used: accuracy for all models and sensitivity and area under curve (AUC) for binary classifier systems.