Predicting students’ final exam scores from course activities can help instructors identify learners who may need additional support and provide feedback based on their performance. Such prediction may also support retention efforts in MOOCs. Peer-assessment activity is one potential source of predictive information: although peer assessments contain uncertainty, they can provide useful signals for modeling final exam performance.

San Diego State University

Google

This paper investigates whether data from peer-based online homework activities in university programming courses (University of Trento, Italy) can predict students' final exam scores. The authors build a linear regression model using features such as students' homework votes, number of tasks completed, and question difficulty level, and evaluate whether the model outperforms random-guess baselines when predicting both continuous exam scores and letter grades (A–F).

Predicting Students’ Final Exam Scores from their Course Activities

The data was gathered from computer programming courses at the University of Trento in Italy and the students were required to participate in the peer-based online homework activities.  The students had to write the questions and the rated their peers answer on the scale 0-5 and the students could also assess their own answers. Final exam score was the number between 18 and 30 . As the authors found, the student votes alone for the homework assignments weren't good predictiors and they decided to use additional features: amount of tasks and the difficulty level of the question. The authors used linear regression model to predict final exam score.  

Building the prediction model (Predicting Students’ Final Exam Scores from their Course Activities)

The authors evaluated their linear regression model by training it on one programming course and testing it on the other. The two cross-course evaluations produced RMSE values of 3.44 and 2.93. To assess whether the model performed better than random guessing, they compared it with baselines that generated exam scores using kernel and uniform distributions. The linear model performed better than these baselines for continuous exam scores. When scores were grouped into letter grades (A–F), however, its advantage over random guessing decreased substantially.

Model Evaluation and Random-Baseline Comparison (Predicting Students’ Final Exam Scores from their Course Activities)

Prior work relevant to predicting final exam scores from course activities falls into two strands. Peer-assessment systems include PRAISE, which supports criteria-based evaluation of programming solutions; PeerWise, in which students create multiple-choice questions for peers to answer; PeerScholar; and Workshop. Earlier student-performance prediction studies often focused on computer science or computer-literacy courses and examined predictors such as current GPA, SAT score, age, mathematics skills, and homework scores.

Learn Before

Related