Context

ShopLens is building a binary classification model to predict whether a customer will purchase within 7 days after viewing a product. A logistic regression baseline was trained on 120,000 sessions, but the team only evaluated it on a single random train/test split and is unsure whether the reported performance is stable enough to trust.

Current Performance

The Problem

The product team wants a more reliable estimate of model performance before launch. Because the positive class is relatively rare and customer behavior varies by traffic source and week, a single split may be giving an overly optimistic or unstable view of performance. You need to explain how you would implement cross-validation for this model evaluation and how you would use the results to decide whether the model is ready.

Requirements

1. Describe which cross-validation strategy you would use and why.
2. Explain how you would prevent data leakage during cross-validation.
3. Specify which metrics you would compute across folds and how you would summarize them.
4. Discuss how you would handle class imbalance when creating folds.
5. Explain how cross-validation results should influence model selection and threshold decisions.

Constraints

• Model training must finish within 2 hours.
• Sessions from the same user should not appear in both train and validation folds.
• The business prefers stable recall over small gains in accuracy.
• The team may later compare logistic regression with gradient boosting using the same evaluation framework.