Abstract:

As students end their bachelor, they are confronted with a new challenge: selecting the set of masters’ courses they will enroll on. Most of the current approaches to this problem tend to recommend courses to students taking into account the topics they have shown interest in the past. Although some approaches make an effort to recommend courses based on students’ potential to get approved on different courses, no approach verifies on which courses students would have better results. We propose the use of one of the most successful techniques in recommendation systems, Singular Value Decomposition (SVD), to capture hidden latent features on historical students grades in order to recommend courses with value to students, given their bachelor performance. We also propose the use of as-soon-as-possible classifiers (ASAP) to enrich student profiling in our recommendation process.

In usual recommendation problems it is used a matrix R, where users are placed in rows and items in columns, and each cell Rij in the matrix corresponds to the rating that the ith user gives to the jth item. SVD is applied in matrix R to compute two different dimensional spaces: the users’ space, that relates users with discovered hidden features, and the items’ space that relates items with the same set of features. The product of these two-dimensional spaces results on an approximation of R, with reduced dimensionality. Our proposal stands on a mapping from the educational context to the usual notation used in recommendation problems: we see students as the users, courses as the items, and the value of the rating corresponds to courses’ grades. To produce recommendations, we first use SVD to compute the students’ and courses’ dimensional spaces. Next, we try to find the most similar historical student to the student being recommended, considering bachelor performance. We then use the features space of the similar student (calculated with SVD) and each of the courses’ feature space to predict the grade of that similar student on each course. Finally, we recommend the N courses with the best-predicted grades.

Our preliminary results show that SVD is effective in predicting students’ grades on masters’ courses. We achieved a mean absolute error 10% smaller than our baseline comparator that sets the predicted grade of each course as the average of the grades achieved on that course. However, these results may be improved using ASAP classifiers to estimate the grades on bachelor courses that students did not get approved until the time of the recommendation, using the grades of bachelor courses where the student got approved. This way, we can recommend masters’ courses to students that have not finished all their bachelor’s courses.

Keywords:

Singular Value Decomposition, Courses Recommendation, Student Profiling, Recommendation Systems.