Predicting students' final performance from participation in on-line discussion forums

• University students' final marks are predicted staring from participation in forums.
• Different data mining approaches are proposed for improving prediction.
• An early prediction before the end of the course is more useful but less accuracy.
• Clustering plus association rules obtain accuracy and highly interpretable models.
• Higher accuracy is obtained when using only the messages related to the subject.

On-line discussion forums constitute communities of people learning from each other, which not only inform the students about their peers' doubts and problems but can also inform instructors about their students' knowledge of the course contents. In fact, nowadays there is increasing interest in the use of discussion forums as an indicator of student performance. In this respect, this paper proposes the use of different data mining approaches for improving prediction of students' final performance starting from participation indicators in both quantitative, qualitative and social network forums. Our objective is to determine how the selection of instances and attributes, the use of different classification algorithms and the date when data is gathered affect the accuracy and comprehensibility of the prediction. A new Moodle's module for gathering forum indicators was developed and different executions were carried out using real data from 114 university students during a first-year course in computer science. A representative set of traditional classification algorithms have been used and compared versus classification via clustering algorithms for predicting whether students will pass or fail the course on the basis of data about their forum usage. The results obtained indicate the suitability of performing both a final prediction at the end of the course and an early prediction before the end of the course; of applying clustering plus class association rules mining instead of traditional classification for obtaining highly interpretable student performance models; and of using a subset of attributes instead of all available attributes, and not all forum messages but only students' messages with content related to the subject of the course for improving classification accuracy.