Scholarlink

One of the key challenges in Massive Open Online Courses (MOOCs) is the high attrition rate. While researchers have worked on predictive models to detect students at risk, one problem is that these models are often trained using data from one course. However, these models face difficulties when generalizing to other contexts. In this line, this work aims to analyze how predictive models can generalize to other similar courses. Particularly, analyses are conducted using a series of three courses about Java programming, with several editions in English and in Spanish, and in teacher-paced and learner-paced modes, and using variables at learner and course level. This way, it is possible to analyze the generalizability considering different combinations of predictive models to forecast MOOC grades and whether or not students pass the course. Results show that it is possible to achieve proper transferability between models in this context, although the predictive power decreases considerably in specific combinations of courses. Moreover, transferability improves when combining two courses. In addition, it is possible to achieve accurate results regardless of the language although significant differences are observed in some cases when transferring models to the courses in a different language. In contrast, there are barely differences when training and predicting using teacher-paced and learner-paced MOOCs, and accurate results are obtained in both cases. These results entail that it is possible to achieve transferable models in MOOCs when using related MOOCs although a drop in the predictive power may appear depending on the course and language.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Analysis of the Generalization of Students’ Success Predictive Models in a Series of Java MOOCs on edX

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Pedro Manuel Moreno-Marcos,
  • Miguel Rodríguez Guillén,
  • Carlos Alario-Hoyos,
  • Pedro J. Muñoz-Merino,
  • Iria Estévez-Ayres,
  • Carlos Delgado Kloos

摘要

One of the key challenges in Massive Open Online Courses (MOOCs) is the high attrition rate. While researchers have worked on predictive models to detect students at risk, one problem is that these models are often trained using data from one course. However, these models face difficulties when generalizing to other contexts. In this line, this work aims to analyze how predictive models can generalize to other similar courses. Particularly, analyses are conducted using a series of three courses about Java programming, with several editions in English and in Spanish, and in teacher-paced and learner-paced modes, and using variables at learner and course level. This way, it is possible to analyze the generalizability considering different combinations of predictive models to forecast MOOC grades and whether or not students pass the course. Results show that it is possible to achieve proper transferability between models in this context, although the predictive power decreases considerably in specific combinations of courses. Moreover, transferability improves when combining two courses. In addition, it is possible to achieve accurate results regardless of the language although significant differences are observed in some cases when transferring models to the courses in a different language. In contrast, there are barely differences when training and predicting using teacher-paced and learner-paced MOOCs, and accurate results are obtained in both cases. These results entail that it is possible to achieve transferable models in MOOCs when using related MOOCs although a drop in the predictive power may appear depending on the course and language.