A comparative study of traditional, deep, and explainable clustering algorithms for behavioral segmentation of MOOC learners

The rapid growth of Massive Open Online Courses (MOOCs) has generated an enormous amount of learner activity data, creating opportunities for personalized learning analytics. However, the complexity, along with the dynamic nature of learners’ behavior, makes the segmentation task challenging. This paper presents a comprehensive traditional evaluation and deep learning-based explainable clustering methods for MOOC learner behavioral segmentation. A large-scale dataset of 50,000 learners, described by behavioral features, was used to benchmark several algorithms, including X-means, agglomerative clustering, DBSCAN, spectral clustering, deep embedded clustering, DEC self-supervised clustering, explainable clustering, k-means with prototype interpretation, and meta-learning clustering. The models were assessed using the silhouette score, the Calinski–Harabasz index, and the Davies–Bouldin index. The results prove that X-Means achieved the best overall performance, with a silhouette of 0.512 and CH of 19,604. However, X-Means achieves a DB of 1.081, outperforming both self-supervised learning approaches and deep learning. Understandable clustering showed strong interpretability for competitive metrics: DB is 1.028, and CH is 23,291, highlighting its potential for actionable insights in educational settings. The DBSCAN performed well, capturing impenetrable structures with a silhouette of 0.458, though at a higher noise fraction. In contrast, spectral clustering exhibited poor separation, with a stability silhouette = –0.325 and DB = 6.096, confirming its limitations for high-dimensional behavioral data. Overall, the results suggest that flexible centroid-based methods, particularly X-means, provide reliable and scalable solutions for segmenting MOOC learners, while enabling more effective personalization through targeted interventions in online education.