Student Engagement Tracking Using Gated Recurrent Unit Networks in Online Educational Environments

Massive Open Online Course (MOOC) and online learning platforms report completion rates that remain persistently below 15% across the industry, a figure that reflects not a failure of content quality but a failure of real-time responsiveness: by the time a learning management system's batch analytics identify a disengaged student, the instructional moment that could have re-engaged them has typically passed. This paper applies a multi-task Gated Recurrent Unit (GRU) network to real-time student engagement tracking, exploiting the GRU's three-gate architecture (reset, update, candidate) to achieve the inference latency required for in-session intervention while retaining the sequence-modelling capacity of recurrent networks. The GRU's three-gate structure yields a theoretical 3:4 parameter ratio relative to an equivalently-sized LSTM, which this study confirms empirically (1.38 million versus 1.84 million parameters; ratio 0.75) and which translates into a 30.5% inference speed advantage (5.84 ms versus 8.4 ms per 30-second window) without sacrificing predictive accuracy. The model is trained and evaluated on a multi-modal corpus of 19,008,000 engagement-labelled 30-second windows (475,200 sessions × 40 windows/session) drawn from 19,800 learners across three MOOC platforms (Platform A: 8,640 learners; Platform B: 6,480; Platform C: 4,680) over 24 months (January 2023–December 2024), combining 28 features across three modalities: facial action units (12 features), clickstream behaviour (10 features), and video player interaction (6 features). A 2-layer GRU (128 units per layer, dropout 0.3) with two output heads—a binary engagement head and a four-class engagement-subtype head (Engaged, Distracted, Passive, Confused)—achieves binary classification accuracy of 88.5%, F1 = 0.872, and AUC = 0.924 on a held-out, student-disjoint test set, outperforming Logistic Regression (AUC 0.724), Random Forest (0.804), and a comparably-sized LSTM (0.884). Deployed as a real-time intervention trigger (nudge dispatched when P(Disengaged) > 0.70 for three consecutive windows) across a matched 9,900-learner treatment group versus a 9,900-learner control group, the system increases module completion from 58.4% to 78.4% (+20.0 percentage points) and reduces dropout from 34.8% to 26.2% (a 24.7% relative reduction), at a measured end-to-end pipeline latency of 1.24 seconds from capture to nudge dispatch.

Hali tarjima qilinmagan