Build and evaluate imbalanced binary classifier

You are given a binary classification dataset with severe class imbalance: the positive rate is about $1\%$.

Each row contains: id, event_date (YYYY-MM-DD), a categorical column region taking values in {NA, EU, APAC, LATAM, MEA}, and numeric features f1 through f50.

The target is $y \in \{0,1\}$.

Tasks: a) Build a reproducible training pipeline that:

• creates a temporal split into train (on or before 2025-06-01), validation (2025-06-02 to 2025-08-01), and test (2025-08-02 to 2025-09-01);
• standardizes the numeric features and one-hot encodes region;
• handles imbalance within cross-validation folds (for example, using class_weight='balanced', or applying SMOTE inside each fold without allowing validation data to leak);
• fits a strong baseline (such as a calibrated logistic regression or gradient boosting model) and returns well-calibrated probabilities (using Platt scaling or isotonic regression on the validation set).

b) Report ROC-AUC and PR-AUC on the test split; also report recall at a 5% false-positive rate and the decision threshold that maximizes F1 subject to recall being at least $0.90$.

c) Explain how you would select the operating point for a production system that must have no more than 2 false positives for every 1,000 predictions.

d) Discuss how calibration could drift over time, and describe one technique for monitoring and re-calibrating without leaking labels.

Overview: This question tests a data scientist's ability to build reproducible machine learning pipelines for imbalanced binary classification, including temporal splitting to prevent leakage, handling class imbalance, feature preprocessing, probability calibration, threshold selection, and monitoring calibration drift.