TLDR: A research paper explores using multivariate time series analysis and deep learning models, specifically Fully Convolutional Networks (FCN), to predict student failure risk in MOOCs. The study, using the OULAD dataset, found that models perform better in STEM courses due to richer interaction data, and that FCN and DOPP are promising for early prediction, highlighting the importance of behavioral data for personalized educational interventions.
Massive Open Online Courses, or MOOCs, have become a cornerstone of democratizing education, offering free and open access to learning for millions worldwide. However, their full potential is often hampered by alarmingly high dropout rates, which can range from 0.7% to 52.1%, with a median of 12.6%. This significant challenge highlights the need for more personalized learning experiences to keep students engaged and successful.
Recent advancements in artificial intelligence and machine learning offer promising solutions to this problem. By analyzing behavioral data from learners, such as clicks and events within the MOOC platform, it’s possible to predict student performance and identify those at risk of failing early on. This proactive approach allows educators to provide timely and tailored feedback, potentially improving completion rates.
A new study, “Prédiction du risque d’échec dans un MOOC : une approche basée sur l’analyse des séries temporelles multivariées,” explores the use of multivariate time series analysis to predict learner failure. The research, conducted by Anass El Ayady, Maxime Devanne, Germain Forestier, and Nour El Mawas, introduces the Fully Convolutional Networks (FCN) model for this purpose, a method rarely applied in this context. It also uniquely analyzes differences in prediction accuracy between STEM (Science, Technology, Engineering, and Mathematics) and SHS (Social Sciences and Humanities) courses, an area often overlooked in previous studies.
The methodology involved transforming student interactions into multivariate time series, where each student’s weekly clicks across various activities (forums, quizzes, resources) formed distinct dimensions. This data was then used to train and evaluate several models, including FCN, DOPP (a model based on LSTM), K-Nearest Neighbors (KNN), and Multilayer Perceptrons (MLP). The study utilized the Open University Learning Analytics Dataset (OULAD), a rich public dataset containing over 10 million clicks from 32,593 students across 22 online courses. For their experiments, the researchers focused on three specific courses: two in STEM (DDD and FFF) and one in SHS (BBB).
The findings indicate that the FCN model, along with DOPP, showed significant promise in predicting learner failure. For binary classification (Pass/Fail), STEM courses (DDD and FFF) consistently yielded higher F1-scores than the SHS course (BBB). The FCN model achieved up to 94% accuracy after 40 weeks in the FFF course. This superior performance in STEM courses is attributed to a higher volume and regularity of interactions, providing richer behavioral data for the models to learn from. In contrast, the SHS course, with its lower interaction density, showed more stagnant performance, especially for shorter observation periods.
For multiclass classification (Distinction, Pass, Fail, Withdrawn), similar trends were observed, though with generally lower F1-scores. Again, FCN and DOPP demonstrated continuous improvement with increased observation time in STEM courses, reaching 66% and 68% respectively for DDD and FFF. This highlights the critical role of extensive temporal data for accurate predictions. Traditional models like KNN, however, showed limited ability to capture complex temporal dynamics and their performance stagnated.
The study concludes that the richness and diversity of weekly interactions are crucial for accurate predictions. Deep learning models, particularly FCN, are effective in capturing complex temporal patterns, outperforming classical approaches. While the study provides valuable insights, it acknowledges limitations, such as relying on a single dataset (OULAD) and a limited number of courses. Future work could explore integrating qualitative data, such as forum discussions, or other behavioral indicators like time spent on activities, to enhance model performance, especially in contexts with limited interaction data. The research also suggests that pedagogical design differences between STEM and SHS courses might influence interaction density and, consequently, predictive model performance.
Also Read:
- Advancing Time Series Prediction with Adaptive Fuzzy Logic and Asymmetric Convolution
- AI Literacy Shapes Student Experience with AI-Powered Assessments, Study Finds
This research contributes significantly to the field of Human Learning Computer Environments (EIAH) by demonstrating the potential of machine learning for personalizing educational pathways. Ultimately, such predictive tools could empower educators with automated analysis systems or dashboards to identify at-risk learners early, enabling targeted pedagogical interventions and fostering greater student success in the evolving landscape of online education. You can read the full paper here.
