Business Context

Data Society wants to predict whether a learner enrolled in a cohort-based training program will complete the course and earn a certificate. The model will be used in Data Society's internal learner success workflow to prioritize outreach for at-risk learners before the final project deadline.

Dataset

You are given a historical dataset of learner enrollments from the last 24 months.

• Rows: 62,000 learner enrollments
• Features: 42 total predictors
• Target: completed_certificate (1 if learner completed the program, 0 otherwise)
• Class balance: 68% completed, 32% did not complete
• Missing data: 12% missing in assessment features for learners who skipped checkpoints; 7% missing in demographics

Success Criteria

A good solution should improve intervention targeting over a naive baseline and support operational use. Aim for ROC-AUC >= 0.82 and F1 >= 0.72 on a held-out test set, while also explaining why the selected model is more appropriate than reasonable alternatives.

Constraints

• Predictions run weekly in batch on the active learner population
• Stakeholders need interpretable drivers of risk, not just raw scores
• Training and scoring should fit within standard Python infrastructure used by Data Society
• Avoid leakage from future engagement or post-deadline features

Deliverables

1. Compare at least three candidate models for this binary classification task.
2. Justify model selection based on data characteristics, bias-variance tradeoff, and operational constraints.
3. Build a reproducible training pipeline with preprocessing and hyperparameter tuning.
4. Evaluate on a held-out test set with appropriate classification metrics.
5. Provide feature importance or coefficient-based interpretation for the final model.