Predicting Academic Performance Based on Learner Traces in a Social Learning Environment

Abstract

Predictive modeling is an important part of learning analytics, whose main objective is to estimate student success, in terms of performance, knowledge, score, or grade. The data used for the predictive model can be either state-based data (e.g., demographics, psychological traits, and past performance) or event-driven data (i.e., based on student activity). The latter can be derived from students’ interactions with educational systems and resources; learning management systems are a widely analyzed data source, while social media-based learning environments are scarcely explored. In this paper, our objective is to predict students’ performance based on their social media traces. Data is collected from a Web Applications Design course, in which students use wiki, blog, and microblogging tools, for communication and collaboration activities in a project-based learning scenario. A total of 343 students, from six consecutive course installments, are included in the study. In addition to the novel settings and performance indicators, an innovative regression algorithm is used for grade prediction. Very good correlation coefficients are obtained and 85% of predictions are within one point of the actual grade, outperforming classic regression algorithms. From a pedagogical perspective, results indicate that, as a general rule, a higher engagement with social media tools correlates with a higher final grade.

Keywords

• Learning analytics
• performance prediction
• large margin nearest neighbor regression
• social learning environment
• social media