Take‑Home: Two End‑to‑End ML Workflows on Tabular Data

Objective

Design and implement two complete machine learning workflows on tabular data (typical of common Kaggle datasets):

• Regression: predict a continuous target.
• Classification: predict a binary or multiclass label.

Assume you have a generic CSV dataset with a mix of numeric and categorical features and a clear target column. If the data are time‑ordered, note time‑series‑specific caveats.

Requirements (for each task)

1. Data cleaning and preprocessing
   • Detect and handle missing values.
   • Detect and handle outliers.
   • Encode categorical features appropriately.
   • Scale features where appropriate.
2. Train/validation/test protocol
   • Random shuffling and splits that avoid leakage.
   • If time‑series or grouped data, use proper split strategies (e.g., forward chaining, GroupKFold).
3. Models
   • A simple baseline (e.g., dummy predictor or regularized linear model).
   • At least one stronger model (e.g., tree‑based, boosted trees).
4. Evaluation
   • Regression: RMSE/MAE (and why).
   • Classification: accuracy, ROC‑AUC, F1 (and why). Use PR‑AUC for heavy class imbalance.
5. Model selection
   • Cross‑validation strategy and hyperparameter tuning.
6. Reproducibility
   • Random seeds, environment pinning, data versioning. Persist splits, models, and configs.
7. Interpretability and reliability
   • Feature importance and partial dependence (or SHAP if available).
   • Calibration checks for classification.
8. Deliverables
   • Pseudocode or code‑level steps for both workflows.
   • Discussion of expected pitfalls and how you would debug underperformance.