End-to-End Binary Classifier Workflow (EDA → Modeling → Fairness → Report)

You are given a labeled tabular dataset and asked to implement a reproducible, end-to-end workflow in Python to analyze the data and train a classifier suitable for deployment.

Assumptions (adapt as needed):

• Input: a CSV file with a binary target column (e.g., target ∈ {0,1}).
• Optional columns: a timestamp column for time-based splits; group columns for fairness checks; an ID column to drop.
• Output: code, metrics, saved model artifact, and a concise text report with a recommended model and expected performance.

Requirements:

1. Data access and schema validation
   • Load data; verify required columns exist; basic type checks and duplicate rows/IDs.
   • Summarize numeric/categorical feature counts and missingness.
2. Exploratory data analysis (EDA)
   • Missing values: counts, percentages, imputation plan.
   • Target leakage checks: suspicious feature names, extremely high target correlation/MI.
   • Class imbalance: distribution and imbalance ratio.
   • Feature distributions: univariate summaries (hist/value counts) and basic outlier flags.
3. Splitting strategy
   • If timestamp present: time-based split (train/validation/test by chronological order).
   • Else: stratified split to preserve class ratio.
4. Baselines
   • Majority-class and simple model baseline (e.g., Logistic Regression with minimal tuning).
5. Preprocessing
   • Numeric: impute (median), scale (standard).
   • Categorical: impute (most frequent), one-hot encode (handle_unknown=ignore); consider rare-category handling.
6. Imbalance handling
   • Use class weights and/or sample weighting; optionally resampling (SMOTE/undersampling) if justified.
7. Model training
   • Train at least two model families (e.g., Logistic Regression and Gradient Boosting).
   • Use cross-validation with hyperparameter tuning (RandomizedSearchCV or equivalent).
8. Metrics aligned to business goal
   • Compute ROC-AUC and PR-AUC; report F1/precision/recall at a chosen threshold.
   • If provided, include cost-sensitive evaluation using FP/FN costs.
9. Calibration (if needed)
   • Assess calibration; calibrate probabilities (Platt or isotonic) if poorly calibrated.
10. Error analysis
    • Confusion matrix and per-slice analysis (e.g., by key categorical/numeric bins).
    • Feature importances (tree-based) and/or coefficients (linear); optionally SHAP.
11. Fairness checks
    • Report metrics by key groups; highlight disparities (e.g., demographic parity, equal opportunity).
12. Reproducibility and report
    • Save the fitted pipeline, metrics JSON, and environment info.
    • Produce a concise recommendation: chosen model, expected performance, and deployment notes.

Deliverables:

• Python code implementing the above.
• Saved model artifact and metrics.
• Short written recommendation with expected performance and guardrails.