Build a leak-free sklearn pipeline
Company: Boston Consulting Group
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Take-home Project
You are training a binary classifier on table data with columns: numeric = [age, balance, n_logins_30d, minutes_watched_7d], categorical = [country, device_type, plan]. Positives are rare (~5%). Write sklearn code to: 1) Split into train/validation with stratification. 2) Create a Pipeline with a ColumnTransformer: numeric → SimpleImputer(strategy='median') + StandardScaler; categorical → OneHotEncoder(handle_unknown='ignore', min_frequency=0.01). 3) Handle imbalance via class_weight='balanced' (and explain when you would instead resample). 4) Tune at least two models (e.g., LogisticRegression and GradientBoostingClassifier or HistGradientBoostingClassifier) using StratifiedKFold and GridSearchCV with scoring='average_precision'. 5) Show that all preprocessing is inside the pipeline so no leakage can occur during CV. 6) Report the best model and its validation metrics (precision, recall, AP) and show how you would persist the trained pipeline for inference. Include code snippets and short justifications for key choices.
Quick Answer: This question evaluates a candidate's ability to construct a leak-free scikit-learn pipeline encompassing column-wise preprocessing, imbalanced binary classification handling, stratified cross-validation, hyperparameter tuning, and model persistence.