Fraud Transaction Detection System

Mohammad Umar

Completed work

Data Analyst

Data Scientist

ML Engineer

pandas

Python

Streamlit

FinTech & Payments

💳 Fraud Transaction Detection - Machine Learning Project

🔹 Author: Mohammad Umar 🔹 Contact: umar.test.49@gmail.com

📌 Section 1: Introduction and Objective

Fraudulent transactions have become a significant concern in the financial industry, especially with the increasing volume of digital payments. This project was developed as part of a machine learning internship to build a system capable of flagging potentially fraudulent transactions.

Client (Assumed):

Financial services firm or payment gateway company needing real-time fraud detection to reduce chargebacks.

🎯 Problem Statement:

Identify whether a transaction is fraudulent (1) or legitimate (0) based on transaction metadata.

❗ Importance:

Prevents financial loss for businesses and customers

Automates fraud investigation workflows

Maintains customer trust and platform integrity

✅ Final Objective:

Build an ML model that:

Accurately classifies transactions

Can be deployed as a user-friendly Streamlit app

Delivers predictions based on user inputs (amount, time, terminal ID, etc.)

📊 Section 2: Dataset

Dataset Overview:

Source: Simulated data based on real fraud patterns

Files: 183 .pkl files (daily transaction logs)

Total Rows: ~1.75 million transactions

Columns: 9 main features + engineered ones

🧾 Important Features:

Feature Description TRANSACTION_ID Unique transaction ID TX_AMOUNT Transaction amount TX_TIME_SECONDS Seconds since midnight TX_FRAUD Target (0: Legit, 1: Fraud) TX_FRAUD_SCENARIO Fraud type ID (if fraud)

⚙ Preprocessing:

Combined all .pkl files into one DataFrame

Converted datetime fields

Removed duplicates and missing values

Engineered temporal features (TX_HOUR, TX_DAY_OF_WEEK)

🔍 Key Observations:

Class imbalance: <1% fraudulent transactions

Certain fraud scenarios dominate minority class

High fraud density during specific hours/weekdays

⚙️ Section 3: Design / Workflow

graph LR
    A[Data Loading] --> B[Cleaning]
    B --> C[EDA]
    C --> D[Feature Engineering]
    D --> E[Train-Test Split]
    E --> F[Model Training]
    F --> G[Evaluation]
    G --> H[Deployment]

🔄 Steps Breakdown:

Data Loading & Cleaning:

Merged 183 transaction files

Handled null/missing values

Converted date fields to datetime format

Exploratory Data Analysis (EDA):

Visualized fraud vs non-fraud distribution

Analyzed transaction amount patterns

Identified peak fraud times (temporal analysis)

Feature Engineering:

Extracted time-based features (hour, day of week)

Removed raw timestamp after feature extraction

Normalized numerical features

Model Training:

Used Logistic Regression with class_weight='balanced'

Stratified 80/20 train-test split (maintained class ratio)

Scaled features using StandardScaler

Evaluation:

Calculated metrics: ✓ Accuracy ✓ Precision ✓ Recall ✓ F1-Score

Generated and analyzed confusion matrix

Plotted ROC curve and precision-recall curve

Deployment:

Built interactive Streamlit web application

Designed intuitive input widgets:

Dropdowns for categorical features

Sliders for numerical values

Date/time pickers

Automated feature transformation pipeline

📈 Section 4: Results

✅ Final Model: Logistic Regression

Metric Score Accuracy 0.7501 Precision 0.0164 Recall 0.4881 F1-Score 0.0317

🧪 Confusion Matrix:

[[261709  86178]
 [  1503   1433]]

📊 Key Insights

Model Performance

High Recall (49%): Catches nearly half of all fraud cases

Low Precision (1.6%): Many false positives due to:

Extreme class imbalance (1:99 ratio)

Intentional recall-precision tradeoff for security

Temporal Patterns

🕒 Peak Fraud Times:

12AM-4AM (overnight window)

Weekends (higher frequency)

Common Fraud Scenarios

Scenario 2: Card-not-present transactions

Scenario 3: Small-amount testing transactions

✅ Section 5: Conclusion

🎯 Achievements

✔ Built complete fraud detection system ✔ 75% accuracy with 49% fraud recall ✔ Production-ready Streamlit interface ✔ Automated feature engineering pipeline

⚠️ Challenges

Class Imbalance:

Only 1% fraudulent transactions

Required careful metric selection

Technical Constraints:

Dataset size limited model complexity

Chose efficient Logistic Regression

Tradeoffs:

Prioritized recall over precision

Accepted false positives for security

🚀 Future Work

Advanced Techniques

Implement SMOTE for class balancing

Test Isolation Forest for anomaly detection

Deployment

Cloud deployment with auto-scaling

Database integration for history

Improvements

Prediction logging system

User feedback mechanism

👨‍💻 Learnings

Handling imbalanced datasets in practice

Balancing theory vs real-world constraints

Importance of model interpretability

End-to-end ML system development

⭐ Star the repository if you found this useful!

Like this project

Completed work

Posted Aug 30, 2025

Developed a fraud detection system using machine learning for financial transactions.

Likes

Views