Select and tune XGBoost hyperparameters

Suppose you are given a binary classification dataset containing 1,000,000 rows and 100 features, of which 20 are numeric and 80 are categorical one-hot encoded, with a 1% positive-class rate. Model training needs to complete within 5 minutes using one 16-core CPU and 32 GB of RAM.

1. Suggest starting XGBoost hyperparameters—eta/learning_rate, max_depth, min_child_weight, subsample, colsample_bytree, lambda, alpha, n_estimators, and max_bin or tree_method—and explain the reasoning for each choice in relation to bias–variance trade-off, class imbalance, and computational limits.

2. Outline a compute-efficient tuning approach, including the search space, early stopping, and a cross-validation design that prevents leakage when the same user appears in more than one fold.

3. Explain precisely how XGBoost treats missing values when splitting trees, and how this behavior interacts with one-hot encoding versus target encoding.

4. With a very small minority class, compare scale_pos_weight, weighted loss, and focal loss; in what situations would each option be most appropriate?

Overview: This question assesses the ability to choose and tune XGBoost hyperparameters, handle severe class imbalance and sparse one-hot encodings, manage missing values, and design compute-efficient training with grouped cross-validation to avoid user-level leakage.