What you write and click reveal : deep learning approaches toward analysis of e-learning forum

Abstract

Massive Open Online Courses (MOOCs) have attracted remarkable attention in the past several years, due to their open access and flexible learning schedules. In addition to integrating conventional coursewares such as lecture videos, reading materials, and quizzes, MOOCs rely on discussion forums to enable socialized learning. A discussion forum, as arguably the only interaction channel in a MOOC, manifests as an interactive platform, where learners can ask questions, share feedback, discuss course-related content, seek help, contribute answers, etc. More importantly, teachers regard discussion forums as the most significant source to monitor the MOOCs, leading to possible course improvement. However, the immense amount of unregulated posts generated by learners with different backgrounds or learning incentives inevitably make a MOOC forum an overwhelming and chaotic environment. A substantial body of literature has been devoted to improving teaching and learning experience in MOOC discussion forums. However, interpreting and monitoring MOOC forums remain challenging due to the following constraints: (1) MOOC discussion forums lack labeled data for supervised machine learning approaches, and most of the related works are based on manual labeling with biased features, implying limited applicability across different MOOCs; (2) Few existing studies targeted precisely at learners, and typical recommendation systems towards learners do not leverage on course settings or learner features; and (3) automatic quality assessment of a vast amount of posts is still an open problem. To tackle these challenges, we propose a systematic framework to analyze and understand the intricacy and dynamics in MOOC discussion forums from three integral perspectives: participants, threads, and posts. First, an improved clustering algorithm is devised to categorize forum participants based on their interactions and performance. Compared with the traditional k-means clustering, our improved approach eliminates outlier effects and avoids local optima triggered by random initialization. The experiment results show that our approach outperforms the baseline in both accuracy and efficiency. Without the necessity of preliminary labels, forum participants can be categorized into distinct clusters, and subsequent analysis is utilized to improve learning experience. Second, we develop a customized recommendation system to suggest relevant threads to forum participants based on the characteristics of courses and learners. Massive threads are embedded into condensed features with topic modeling, and a composite neural network is designed for thread recommendation. The overall accuracy of our recommender shows promising results on distinct MOOCs. Last but not least, we employ a convolutional neural network (CNN) for automatic quality assessment of forum posts. The composite features are synthesized from text-based content, stylistic features, and video-watching records. Our customized model demonstrates a decent performance in predicting endorsed posts, helping both teachers and learners identify high-quality content efficiently. In conclusion, our systematic framework towards MOOC forum analysis can assist both teachers and learners in deriving high values from forums and thereby optimizing online learning experience. Future work includes analysis of silent learners based on clickstream data, text mining in noisy context, and implementation with user-friendly tools.