Analysing Student Performance using Sparse Data of Core Bachelor Courses

Abstract

Curricula for Computer Science (CS) degrees are characterized by the strong occupational orientation of the discipline. In the BSc degree structure, with clearly separate CS core studies, the learning skills for these and other required courses may vary a lot, which is shown in students’ overall performance. To analyze this situation, we apply nonstandard educational data mining techniques on a preprocessed log file of the passed courses. The joint variation in the course grades is studied through correlation analysis while intrinsic groups of students are created and analyzed using a robust clustering technique. Since not all students attended all courses, there is a nonstructured sparsity pattern to cope with. Finally, multilayer perceptron neural network with cross-validation based generalization assurance is trained and analyzed using analytic mean sensitivity to explain the nonlinear regression model constructed. Local (withinmethods) and global (between-methods) triangulation of different analysis methods is argued to improve the technical soundness of the presented approaches, giving more confidence to our final conclusion that general learning capabilities predict the students’ success better than specific IT skills learned as part of the core studies.