Credit Scoring Model and MLOps with MLflow

Python, Machine Learning, MLOps, MLflow, LightGBM, XGBoost, Scikit-learn, Credit Scoring, Optuna, Feature Engineering

Credit Scoring Model and MLOps with MLflow

Context

As part of an MLOps-oriented Data Science project, I developed a Credit Scoring model to estimate a customer's risk of loan default based on financial and behavioral data. The project covers the full model lifecycle, from data preparation to experiment tracking and version management with MLflow.

Objectives

• Build a classification model for payment default risk.
• Handle a highly imbalanced dataset.
• Optimize the decision threshold according to a business cost.
• Compare several Machine Learning algorithms.
• Implement an MLOps approach with MLflow.
• Ensure traceability of experiments and models.

Architecture

Data Sources
(Home Credit)
        │
        ▼
Merging & Cleaning
        │
        ▼
Feature Engineering
        │
        ▼
Model Training
        │
        ▼
MLflow Tracking
        │
        ▼
Model Registry
        │
        ▼
MLflow Model Serving

Technologies Used

• Python
• Pandas
• NumPy
• Scikit-learn
• LightGBM
• XGBoost
• MLflow
• Optuna / GridSearchCV
• Matplotlib
• Jupyter Notebook

Achievements

Data Preparation

• Merging multiple datasets.
• Cleaning and handling missing values.
• Encoding categorical variables.
• Creating new variables (Feature Engineering).
• Analyzing class imbalance.

Modeling

Comparison of several classification models:

• Logistic Regression
• Random Forest
• LightGBM
• XGBoost
• Multi-Layer Perceptron (MLP)

Performance evaluation through stratified cross-validation.

Optimization

• Hyperparameter search with GridSearchCV or Optuna.
• Decision threshold optimization.
• Definition of a business cost function giving more weight to false negatives.
• Model comparison based on business and technical metrics.

MLOps

Setting up an initial MLOps chain with MLflow:

• experiment tracking
• automatic logging of parameters
• metric tracking
• artifact storage
• model version management (Model Registry)
• testing Model Serving

Evaluation

Model comparison using several indicators:

• AUC-ROC
• Recall
• Precision
• F1-Score
• Accuracy
• Custom business cost

Skills Demonstrated

Data Science

• Data preparation
• Feature Engineering
• Exploratory analysis (EDA)
• Handling imbalanced data

Machine Learning

• Supervised classification
• Cross-validation
• Hyperparameter optimization
• Decision threshold optimization
• Business-driven model evaluation

MLOps

• MLflow Tracking
• Experiment Tracking
• Model Registry
• Model Serving
• Model version management
• Experiment reproducibility

Python Development

• Pandas
• Scikit-learn
• LightGBM
• XGBoost
• Training pipelines

Business Analysis

• Credit Scoring
• Risk analysis
• Business cost function optimization
• Supporting financial decision-making

Outcome

This project allowed me to develop a robust Credit Scoring model while implementing initial MLOps practices. Using MLflow ensured full traceability of experiments, model version management, and readiness for future industrialization.

Skills Acquired

• Python
• Pandas
• NumPy
• Scikit-learn
• Machine Learning
• Classification
• Credit Scoring
• LightGBM
• XGBoost
• MLflow
• MLOps
• Experiment Tracking
• Model Registry
• Model Serving
• Feature Engineering
• Cross Validation
• GridSearchCV
• Optuna
• Data Visualization
• Jupyter Notebook