Video-based instruction plays an increasingly prominent role in higher education, but little is known about how students actually engage with such materials across an academic cycle. This study analyses one year of learning analytics from an engineering YouTube channel integrated into university teaching to examine temporal and design-related patterns of student engagement. A time series analysis established how and when students interact with the channel. Dominant weekly (approx. 7-day) and shorter (approx. 3-day) viewing cycles were identified, with engagement intensifying sharply ahead of assessment periods. This temporal context motivated a closer look at which videos students return to. Principal Component Analysis (PCA) was applied to reduce dimensionality and create a composite behavioural performance score. This score served as the basis for a clustering analysis that identified three distinct video categories: high-retention/low-popularity, balanced, and high-popularity/low-retention videos. Video duration and playlist position emerged as key characteristics separating these clusters. Building on this, a predictive model was trained exclusively on video pre-upload features to forecast the composite score. Both Random Forest (RF) and Linear Regression (LR) models demonstrated strong predictive power on an unseen testing set, achieving R 2 values of 0.796 and 0.810, respectively. The results were further interpreted through Cognitive Load Theory and Mayer's Multimedia Learning principles. While findings are scoped to behavioural platform metrics and no direct learning outcome data were incorporated, the results provide practical guidance for future video design and identify the pre-upload characteristics that most influence how students interact with educational video content.
Soares et al. (Mon,) studied this question.