Credit-Risk-Prediction-using-Explainable-AI

in this project we predict the likelihood of credit default , while applying explainable AI techniques like LIME and SHAP for transparent decision-making

📁 Table of Contents

• Project Overview
• Data Description
• EDA Summary
• Data Preprocessing
• Dimensionality Reduction
• Model Training Overview
• Model Explainability
• Results & Insights
• Installation

---

Project Overview

This project focuses on predicting the likelihood of a customer defaulting on a loan by leveraging historical credit data, with a strong emphasis on explainability and fairness. Traditional credit scoring models often function as black boxes, leading to challenges in transparency and accountability.

To address this, we employ interpretable machine learning techniques such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) to provide clear insights into model decisions. Furthermore, the project focuses to mitigate potential biases related to sensitive attributes like gender or age.

By balancing predictive performance with interpretability and ethical considerations, the project aims to support responsible decision-making in credit risk assessment.

---

Data Description

The Default of Credit Card Clients Dataset is sourced from the UCI Machine Learning Repository and contains information on the credit history and behavior of 30,000 credit card holders in Taiwan.

---

Dataset Overview

• Source: UCI Machine Learning Repository
• Original Provider: Taiwan Economic Journal
• Published By: Yeh, I.C., & Lien, C.H. (2009)
• Number of Instances: 30,000
• Number of Features: 23
• Prediction Target: default.payment.next.month (1 = default, 0 = no default)

---

Feature Details

Column Name	Description
ID	Unique ID for each client
LIMIT_BAL	Amount of given credit (NT dollar): includes individual and family credit
SEX	Gender (1 = male, 2 = female)
EDUCATION	Education level (1 = graduate school, 2 = university, 3 = high school, etc.)
MARRIAGE	Marital status (1 = married, 2 = single, 3 = others)
AGE	Age in years
PAY_0 to PAY_6	Repayment status in months from September to April (0 = paid, 1 = delay 1 month, ..., -1 = pay duly)
BILL_AMT1 to BILL_AMT6	Bill statement amount (NT dollar) from September to April
PAY_AMT1 to PAY_AMT6	Amount paid in previous months (NT dollar) from September to April
default.payment.next.month	Target variable (1 = default, 0 = no default)

---

Key Characteristics

• Multivariate: Numeric + categorical features
• Supervised Classification: Predict default (yes/no)
• Real-world Financial Data
• Moderate class imbalance: ~22% defaulted clients

Download

• Direct Dataset Link (CSV)

---

EDA Summary

1. Class Distribution

The dataset exhibits a significant class imbalance — only 20% of records belong to the default class, while the remaining 80% are non-default. This imbalance can negatively impact model performance, especially for minority class predictions. To address this, we later apply resampling techniques like SMOTE to balance the classes.

---

2. Categorical Feature Distributions

The data reveals several interesting patterns:

• Gender: The default rate is noticeably higher among male clients than females.
• Education Level: Higher education levels appear to correlate with a lower risk of default.
• Marital Status: Married individuals show a slightly elevated risk of default compared to single clients.

However, these observations may be coincidental and not necessarily indicative of true associations. To verify the statistical significance of these relationships, we conduct a Chi-Square Test of Independence.

Chi-Square Test of Independence

Hypotheses:

• Null Hypothesis (H₀):
  There is no association between the categorical variable (e.g., education level) and loan default status; any variation is due to chance.

• Alternative Hypothesis (H₁):
  There is a significant association between the categorical variable and default status; observed differences are not random.

The test results help determine which categorical features are statistically relevant for predicting credit risk.

---

3. Payment Status Distribution

Boxplots and count plots for monthly payment status (PAY_0 to PAY_6) highlight important trends:

• Clients with delays of one month or less are much less likely to default.
• Among all months, September (PAY_1) demonstrates the highest discriminative power in identifying defaulters.

This makes PAY_1 a key feature in predicting credit default risk and justifies its priority in model training and feature selection.

4. Correlation Analysis

There is a strong positive correlation between:

• Billing amounts (BILL_AMTn)
• Payment statuses (PAY_n)

This may suggest redundant information across these features. As a result, PCA (Principal Component Analysis) or feature selection techniques may be applied to reduce dimensionality and multicollinearity.

---

5. Credit Limit vs. Default

Although defaulters tend to have a lower mean credit limit, the distributions of credit limit for defaulters and non-defaulters overlap significantly. Therefore:

• Credit Limit alone is not a strong predictor.
• However, it may still be useful when combined with other features.

---

6. Age Distribution

The age range of 30–40 shows a higher density of defaulters compared to other age groups. This indicates a potential risk band that could enhance model performance when captured properly.

---

7. Monthly Payment Amounts

Monthly Payment Amount and Default Status

Boxplots of monthly payment amounts (PAY_AMT1 to PAY_AMT6) indicate:

• Non-defaulters typically make higher payments across all months.
• Median payment amounts are consistently lower among defaulters.
• There are many outliers, but the overall central tendency makes payment amount a useful predictor.

---

---

Data Preprocessing

Encoding & Splitting

• Applied one-hot encoding to convert categorical variables.
• Used a 1:4 stratified train-test split to maintain class distribution.

Feature Scaling

While tree-based models are scale-invariant, others benefit from scaling. Two methods were considered:

• Normalization (Min-Max Scaling): $$ X_{norm} = \frac{X - X_{min}}{X_{max} - X_{min}} $$
• Standardization (Z-score Scaling): $$ Z = \frac{X - \mu}{\sigma} $$

✅ Final choice: Normalization, due to better handling of outliers.
📌 Scaling is applied only on training data statistics to prevent leakage.

Handling Class Imbalances

Imbalanced datasets can cause models to become biased toward the majority class, leading to poor performance on the minority class. To address this, we applied the following resampling techniques:

🔹 Cluster Centroid Undersampling (Undersampling)

This technique reduces the number of majority class samples by clustering them using KMeans and replacing them with the cluster centroids.

• ✅ Reduces class imbalance by shrinking the majority class size.
• ✅ Retains the structural distribution of the original data.
• 📌 Best used when the majority class is significantly overrepresented or contains redundant points.

🔹 Simple SMOTE (Oversampling)

SMOTE creates new synthetic instances of the minority class by interpolating between a data point and one of its k nearest neighbors.

• ✅ Increases the size of the minority class to balance the dataset.
• ✅ Prevents overfitting caused by simple duplication.
• 📌 Effective when the minority class is well-distributed and not clustered.

🔹 KMeans-SMOTE (Oversampling)

This hybrid method first clusters the minority class using KMeans, then applies SMOTE within each cluster to generate synthetic samples.

• ✅ Produces more localized and realistic synthetic points.
• ✅ Captures internal structure of the minority class more effectively.
• 📌 Useful when the minority class has subgroups or complex distributions.

---

Note: We evaluated model performance using each resampling strategy to identify the most effective technique for improving minority class prediction. The final choice was based on validation metrics like F1-score and recall.

---

Dimensionality Reduction: Interpretability vs Performance

While building our credit risk models, we evaluated two approaches for reducing the feature space, each with its own trade-off between performance and interpretability.

---

Path 1: Feature Selection + SHAP (High Interpretability)

• We used methods like SelectKBest and RFE to retain original, domain-relevant features.
• These selected features were passed to models such as Logistic Regression and XGBoost.
• With this setup, SHAP values directly mapped to the actual features, offering clear, regulator-friendly insights.

✅ This approach worked best when transparency and explainability were our primary concerns.

---

Path 2: PCA + SHAP

• Applied Principal Component Analysis (PCA) to reduce dimensionality on:

  featuresToReduce = [f'PAY_{i}' for i in range(1, 7)] + \
                     [f'BILL_AMT{i}' for i in range(1, 7)] + \
                     [f'PAY_AMT{i}' for i in range(1, 7)]

• The top 7 principal components explained approximately 95% of the variance in the original feature set.

---

Summary of Our Findings

Feature Selection + SHAP

• ✅ Interpretability: High — SHAP values directly align with original features.
• ⚪ Model Performance: Reasonable, with good transparency.
• ✅ SHAP Clarity: Immediate and intuitive.
• 📌 Best Used When: Regulatory compliance, stakeholder explanations, or early-stage data exploration are important.

PCA + SHAP

• ⚠️ Interpretability: Lower — components are harder to explain.
• ✅ Model Performance: Often higher due to variance capture.
• ⚠️ SHAP Clarity: Requires reverse mapping for explanation.
• 📌 Best Used When: Predictive accuracy is key, and some interpretability can be traded off.

---

📌 What We Chose: For generating clear, actionable model explanations using SHAP and LIME, we prioritized feature selection over PCA to maintain interpretability without sacrificing too much performance.

---

Model Training Overview

Models Trained

• Logistic Regression
• Decision Tree
• Random Forest
• XGBoost
• Kernel Approximation SVM (RBFSampler + LinearSVC)

ℹ️ Why Kernel Approximation SVM?
Standard SVM with non-linear kernels (like RBF) is computationally expensive on large datasets.
Kernel Approximation SVM uses RBFSampler to approximate the RBF kernel, enabling the use of a faster LinearSVC, significantly improving scalability.

Feature Selection

• Performed using SelectFromModel with a RandomForestClassifier.
• Retains only the most important features based on impurity-based importance.

Hyperparameter Tuning

• Conducted using GridSearchCV with 3-fold cross-validation.
• F1 Score used as the scoring metric to balance precision and recall due to class imbalance.
• No tuning performed for Kernel Approximation SVM (used fixed configuration).

Evaluation Metrics

• Accuracy
• Precision
• Recall
• F1 Score

Why F1 Score?
F1 Score was prioritized to handle class imbalance by balancing false positives and false negatives.

---

Model Performance Visualization

Performance of each model across different class balancing techniques:

• Logistic Regression
  

• Decision Tree
  

• Random Forest
  

• XGBoost
  

• Kernel Approximation SVM
  

---

---

Model Explainability

To ensure transparent, trustworthy decision-making in credit risk assessment, we used two widely adopted explainable AI (XAI) techniques: SHAP and LIME. These tools help break down complex model behavior and explain individual and global predictions.

---

1. 🔍 SHAP (SHapley Additive exPlanations)

SHAP explains the model's predictions by assigning each feature a contribution value for every prediction. These values are based on game-theoretic Shapley values.

🧮 Global Feature Importance

• This bar chart shows mean absolute SHAP values, which reflect how much each feature contributes to predictions across the dataset.
• Key insights:
  • PAY_1, LIMIT_BAL, and PAY_AMT1 are the most influential features.
  • Features related to payment history (PAY_1–PAY_6) and credit usage (LIMIT_BAL, BILL_AMTn) dominate the importance list -model is not based on any biasis based on gender and education

---

2. 🧩 LIME (Local Interpretable Model-Agnostic Explanations)

While SHAP offers both global and local explanations, LIME specializes in local interpretability, helping to understand individual decisions.

🧑 Local Explanation Example

• This LIME explanation shows how each feature affects a single customer's prediction.
• Features pushing the model toward default are shown in red, and those pushing toward non-default are in green.
• This helps stakeholders and users understand why a specific prediction was made, especially in sensitive financial contexts.

---

3. 💧 SHAP Waterfall Plot: Individual Prediction Breakdown (XGBoost)

We also use SHAP waterfall plots to trace how an individual prediction is formed from base value to final output.

📉 SHAP Waterfall Plot

• This visualization starts at the model’s base value (average prediction).
• Each feature adds or subtracts from this baseline, showing their directional impact.
• For example:
  • PAY_1 might increase the likelihood of default significantly.
  • PAY_AMT1 or LIMIT_BAL might reduce it.

---

✅ Summary

Technique	Purpose	Advantage
SHAP	Global + Local Explanation	Theoretically sound, interpretable for both data-wide and individual predictions
LIME	Local Explanation	Model-agnostic, helpful for user-level insights
Waterfall Plot	Per-instance decision breakdown	Visualizes how each feature moves the prediction

These tools enable our model to be interpretable, fair, and regulatory-compliant, making it suitable for deployment in real-world financial applications.

---

Installation

Follow the steps below to set up the project environment and run the application:

1. Clone the Repository

git clone https://github.com/divyanshjain122/Credit-Risk-Prediction-using-Explainable-AI.git 

2. Install Dependencies

Navigate into the project directory and install the required Python packages:

cd Credit-Risk-Prediction-using-Explainable-AI
pip install -r requirements.txt 

3. Run the Streamlit Application

Inside the project folder, launch the Streamlit app:

streamlit run src/app.py