Machine Learning Project

Credit Card
Fraud Detection

Catching the 0.17% of transactions that cost billions

91% PR AUC
284,807 Transactions
492 Fraud Cases
7 SOTA Models
+3.3% Improvement
💡 Click any graph to expand • Click again to close

Complete ML Results

12 models tested. One winner.

0.173% Fraud Rate
88.49% Original Score
91.04% After Augmentation

Data Exploration

Understanding the patterns

Class Distribution
!

Extreme Imbalance

Only 492 frauds in 284,807 transactions

Amount Analysis
$

Amount Patterns

Fraudsters prefer smaller transactions

Time Analysis
T

Temporal Patterns

Fraud timing throughout the day

Feature Correlation
~

Feature Correlations

V14, V17, V12 most correlated with fraud

🏆

XGBoost Wins

Gradient boosting with data augmentation crushed 11 other models including deep learning.

91.04%
Augmented XGBoost
vs
73.61%
Deep Neural Network

Model Analysis

Understanding what drives the predictions

Feature Importance
*

Feature Importance

V14 + V17 account for 34% of decisions

Threshold Analysis
=

Threshold Trade-offs

Precision vs recall business impact

PR Curve
^

PR Curve

Area under curve: 91.04%

Model Comparison
|

All Models Ranked

Gradient boosting dominates

SHAP Feature Importance
S

SHAP Explainability

Understanding why the model makes decisions

All Models PR Curves
~

SOTA Model Comparison

XGBoost, CatBoost, LightGBM, TabNet, Transformer

Summary Dashboard

The complete picture