Abstract:

Test results are key determinants of student performance. More specifically, intermediate and final tests are popular and powerful tools that teachers can use to track their students' learning progress and help them improve their performance at the same time. Knowing this, we questioned whether there was unexplored potential information that could be extracted from student tests aside from test results.

In this work, we are interested in finding out how complex each test appeared to the professor and the students and figuring out whether their perceptions of complexity aligned. To accomplish this, we use 251 test scores obtained in 8 tests by 42 students attending a structured programming course at Brainster Next College. In addition, upon making the test available for the students, the professor was asked to rate how difficult each test was. The same question was posed to each student as they uploaded their solutions. Both the professor and the students were unaware of what the other had selected.

On this data, we used the K-means clustering algorithm to check whether there are discernible patterns in the relationship between the differences between professor and student perceived levels of test complexity and the number of points students earned on these tests. The resulting clusters were analyzed using statistical methods like T-distributed stochastic neighbor embedding (t-SNE).

Our results show that students who perceive a test as more difficult than the professor tend to achieve worse results than those who perceive the test to be easier. While our analysis is only exploratory, the results are encouraging as they hint at the possibility of finding out early on which students will do well in a subject and which will struggle.

Keywords:

Student performance, student performance prediction, exam complexity, learning management systems, k-means clustering, t-SNE.