💳 UPI Fraud Detection Using Machine Learning 📌 Overview

Unified Payments Interface (UPI) is a real-time digital payment system widely used in India. With the rapid growth of digital transactions, detecting fraudulent activities has become increasingly important.

This project focuses on building a machine learning model to detect fraudulent UPI transactions using classification techniques.

🎯 Objective

To develop a robust fraud detection system that can accurately classify transactions as:

✅ Legitimate (0)

🚨 Fraudulent (1)

📊 Dataset & Preprocessing

The following steps were performed:

Removed duplicate records

Checked and handled missing values

Converted date column to datetime format

Extracted time-based features (Year, Month)

Applied One-Hot Encoding for categorical variables

Removed unnecessary ID-based columns

Handled class imbalance using SMOTE (Synthetic Minority Over-sampling Technique)

🤖 Models Implemented

The following machine learning models were trained and compared:

Decision Tree

Random Forest

Gradient Boosting

XGBoost

After evaluation and hyperparameter tuning, Random Forest performed the best and was selected as the final model.

⚙️ Model Optimization

Applied SMOTE to balance the training dataset

Performed Hyperparameter Tuning using GridSearchCV

Re-trained Random Forest using optimal hyperparameters

Although RandomizedSearchCV is more computationally efficient, GridSearchCV was used for exhaustive parameter tuning.

📈 Final Model Performance (Random Forest)

Accuracy: 94.9%

ROC-AUC Score: 94.9%

F1 Score: ~0.95

These results indicate strong classification performance and effective fraud detection capability.

🧠 Key Insights

Random Forest performed best due to its ensemble learning approach and ability to reduce overfitting by combining multiple decision trees.

Handling class imbalance using SMOTE significantly improved recall and fraud detection capability.

ROC-AUC and F1-score were prioritized over accuracy due to the nature of fraud detection problems.

🛠 Technologies Used

Python

Pandas

NumPy

Scikit-learn

Imbalanced-learn (SMOTE)

Matplotlib / Seaborn