Disengagement prediction estimates the likelihood that a learner will abandon a session before it ends, using behavioural pattern changes observed in real time.
What it is
Models are typically trained by retroactive labelling: when a confirmed abandonment occurs, the signals from the preceding minutes are labelled as pre-disengagement. Over many sessions this builds a statistical picture of what leaving looks like before it happens.
Features usually include interaction latency, dwell against baseline, tab-switch duration and scroll velocity.
Why it matters
Boote, Agarwal and Mostow reported 73.3% accuracy with 40% of the activity remaining. In a forty-minute module that is roughly sixteen minutes of warning.
A common confusion
73% accuracy means more than a quarter of flags are wrong. That is fine for adapting content invisibly and unacceptable for anything touching an individual's record.
Related
See also at-risk learner, adaptive learning and cyberloafing. See the full glossary for the rest, or identifying disengaged learners early for the wider context.
See where your content loses people
Book a walkthrough and we will show you the engagement data on your own content.