How tergelitu/ml-models-for-credit-scoring-benchmark works
Project Type & Purpose
• This is an ML benchmarking pipeline (not a multi-service web app). It’s packaged as a Jupyter notebook (result.ipynb) that orchestrates the end-to-end credit-scoring workflow: data ingestion → preprocessing → feature engineering → model training & tuning → evaluation & visualization.
Key Files & Directories
• result.ipynb: orchestration of all steps
• requirements.txt: Python libraries (pandas, scikit-learn, XGBoost, LightGBM, CatBoost, matplotlib, seaborn, etc.)
• data/ (mentioned in README): raw CSVs from Kaggle
• outcome/: generated plots, flowcharts, ROC curves, 3D visualizations
Main Components
a. Data Ingestion
• External service: Kaggle dataset (downloaded to local data/ folder)
• Component: Data Loader (pandas)
b. Data Preprocessing & Cleaning
• Null handling, outlier removal, encoding
• Tools: pandas, NumPy
c. Feature Engineering
• Aggregation of auxiliary tables, custom ratios, time-based features
• Output: single merged DataFrame keyed by SK_ID_CURR
d. Sampling & Feature Selection
• Stratified sampling
• Feature importance ranking (RF, XGBoost, LGBM, CatBoost, GB)
• Correlation filtering, leakage detection
e. Model Training & Evaluation
• Parallel training of multiple models
• ROC AUC computation & comparison
• Visualization: ROC curves
f. Hyperparameter Tuning
• GridSearchCV (scikit-learn) or built-in tuning of LightGBM/CatBoost/GB
g. Score Binning & 3D Intersection
• K-means binning into deciles
• 3D plot of bin intersections across three model probabilities
h. Acceptance Region Analysis
• Contiguous region with bad-rate threshold
• 3D cutoff area visualization
i. Outputs
• Plots and flowcharts saved in outcome/
Relationships & Data Flow
• Data flows linearly from ingestion → preprocessing → feature engineering → sampling → training → evaluation → visualization → storage.
• Hyperparameter tuning loops back into the training component.
• Visual outputs feed into the outcome folder.
Architectural Patterns & Principles
• Pipeline pattern: each stage transforms the dataset, outputs feed into next stage.
• Modular stages (each can be extracted into functions/modules if codebase grows).
• Single orchestrator (notebook) – consider separating into scripts or a DAG for production.
Technologies & Frameworks
• Python ecosystem: pandas, NumPy for ETL; scikit-learn for modeling & grid search; XGBoost, LightGBM, CatBoost for specialized boosting; matplotlib/Seaborn for plots; Jupyter for orchestration.