Deep Learning for Discussion-Based Cross-Domain Performance Prediction of MOOC Learners Grouped by Language on...

Area of Science:

• Educational Technology
• Machine Learning in Education
• Natural Language Processing

Background:

• Learner behavior analysis in MOOCs traditionally uses numerical activity data.
• Machine learning models effectively predict learner performance using these numerical features.
• There's a need to explore richer data sources like learner comments for enhanced prediction.

Purpose of the Study:

• To investigate the predictive power of incorporating learner comment text data, categorized by English language proficiency, into MOOC performance prediction models.
• To compare the effectiveness of models using only comment data versus those using a combination of comment and traditional activity features.
• To assess the generalizability of these learning algorithms across different MOOCs.

Main Methods:

• Collected and analyzed over 420,000 comments from FutureLearn language-focused MOOCs.
• Developed a method to infer learners' first language based on their country.
• Trained and compared machine learning models, including Bidirectional LSTM, using extracted numerical features, comment text, and a combination of both.

Main Results:

• Models using only comment text data showed weaker predictive performance compared to models incorporating extracted activity features.
• A combined approach using both comment content and extracted activity features yielded improved prediction of learner engagement and success.
• The Bidirectional LSTM model, trained on discussion comments, achieved 73% success in predicting learner performance on a different MOOC, demonstrating potential despite variations across models.

Conclusions:

• Incorporating learner comment analysis, alongside traditional activity metrics, enhances the prediction of learner outcomes in MOOCs.
• Learner language proficiency, inferred from comments and country data, can be a valuable feature in educational data mining.
• While performance varies, the developed methods show promise for improving the generalizability and accuracy of learning analytics across diverse MOOCs.