Bagging vs Boosting for Integrity Risk

Business Context

Meta's Integrity team wants a model to predict whether a newly created Facebook account will be actioned for coordinated spam or fake engagement within 7 days. You need to compare bagging and boosting in a realistic classification setting and recommend which ensemble approach should be deployed.

Dataset

You are given a training table built from account creation, graph, and early activity signals.

• Size: 420K accounts, 42 features
• Target: enforced_7d — 1 if the account is actioned within 7 days, else 0
• Class balance: 6.4% positive, 93.6% negative
• Missing data: ~12% missing in graph features for cold-start accounts; ~4% missing in device-level fields

Success Criteria

A good solution should:

• Beat a single decision tree baseline by a meaningful margin
• Show a clear comparison between a bagging method and a boosting method
• Achieve PR-AUC >= 0.42 and recall >= 0.75 at precision >= 0.30 on the test set
• Explain which method is preferable for Meta's production constraints

Constraints

• Batch scoring runs every 15 minutes; p95 inference latency should stay under 50 ms per 1K accounts
• The model must support feature importance analysis for Integrity analysts
• Retraining happens weekly; the pipeline should tolerate moderate feature drift

Deliverables

1. Train a single decision tree baseline, a bagging model, and a boosting model.
2. Explain the conceptual difference between bagging and boosting in the context of this dataset.
3. Compare performance using PR-AUC, ROC-AUC, recall at fixed precision, and calibration.
4. Recommend one approach for deployment and justify the tradeoffs.
5. Describe how you would monitor degradation after launch.