Lernnavi — Disengagement Prediction
Predicting disengagement on a Swiss learning platform.
Lernnavi has no course end date or dropout label. The task is to distinguish sustained inactivity from school holidays.
Method
A two-stage pipeline predicts whether learners return after three weeks, then whether they enter an interruption of at least four weeks. Logistic regression, LightGBM and LSTM models are compared using twelve-week sequences, ablations and fairness analysis.
Results
The best Stage 2 model achieves an AUC-ROC of 0.9125, as reported in the README. Calendar features are important. Differences across cantons require careful interpretation; predictions are intended to support human decisions.