Gaze cluster analysis reveals heterogeneity in attention allocation and predicts learning outcomes

Instructional videos need to maintain learners' attention to foster learning, therefore, a fine-grained measurement of attention is required. Existing gaze measures like inter-subject correlation (ISC) assume a singular focal point deemed meaningful for indicating attention. We argue that multiple meaningful foci can exist and propose an automatically generated gaze measure labeled gaze cluster membership (GCM). By applying the density-based clustering in spatial databases (DBSCAN) algorithm to gaze position data from over 100 participants, we categorize viewers as attentive when they are part of a cluster and as inattentive when they are not. Using two videos, we demonstrate that our settings of DBSCAN generate meaningful clusters. We show that low ISC values (neuronal and eye tracking data) during multiple meaningful foci do not necessarily indicate a lack of attention. Additionally, GCM predicts participants' self-reported mental effort and their tested knowledge. Our innovative approach is of high value for assessing learner attention and designing instructional videos.