🕵️ Fraud Transaction Detection

https://fraudtransactiondetection-uee8kzouwxas9pc3rve5mc.streamlit.app/

This is an end-to-end machine learning project to detect fraudulent financial transactions using a stacking ensemble model (Random Forest + XGBoost). It includes preprocessing, modeling, evaluation, prediction, and a user-friendly Streamlit web app.

Key Features

• Real-time transaction risk evaluation
• Multi-model approach for improved accuracy
• Comprehensive feature engineering pipeline
• Scalable architecture for production deployment
• Detailed performance metrics and monitoring

Architecture Overview

The system employs a modular design with clear separation of concerns, ensuring maintainability and scalability.

flowchart TD
    subgraph Input["Input Layer"]
        A[Transaction Data]
    end

    subgraph Processing["Processing Pipeline"]
        B[Feature Engineering]
        C[Model Selection]
        D[Prediction Engine]
    end

    subgraph Models["ML Models"]
        E[XGBoost Model]
        F[KNN Model]
        G[SVM Model]
    end

    subgraph Output["Results"]
        H[Fraud Score]
        I[Alert System]
    end

    A --> B
    B --> C
    C --> D
    E --> D
    F --> D
    G --> D
    D --> H
    H --> I

    classDef input fill:#93c47d,stroke:#6aa84f,color:#000
    classDef process fill:#6fa8dc,stroke:#3d85c6,color:#000
    classDef model fill:#e06666,stroke:#cc0000,color:#000
    classDef result fill:#ffd966,stroke:#bf9000,color:#000

    class A input
    class B,C,D process
    class E,F,G model
    class H,I result

Loading

The architecture diagram above illustrates the system's four main layers:

• Green: Input layer handling raw transaction data
• Blue: Processing pipeline managing feature engineering and model selection
• Red: ML models (XGBoost, KNN, SVM) contributing to predictions
• Yellow: Results layer generating fraud scores and triggering alerts

Each transaction flows through the feature engineering pipeline before being evaluated by multiple machine learning models. The Prediction Engine combines outputs from all models to generate a final fraud score, which triggers alerts based on configurable thresholds.

Technical Implementation

Core Components

fraud_transaction_detection/
├── src/
│   ├── features/
│   │   └── feature_engineering.py
│   ├── models/
│   │   ├── xgboost_model.py
│   │   ├── knn_model.py
│   │   └── svm_model.py
│   ├── pipeline/
│   │   └── prediction_pipeline.py
│   └── utils/
│       └── metrics.py
└── notebooks/
    └── analysis.ipynb

Model Comparison

Model	Balanced Accuracy	Precision	Recall	F1 Score
XGBoost	0.881 ± 0.017	0.963 ± 0.007	0.763 ± 0.035	0.851 ± 0.023
KNN	0.705 ± 0.037	0.942 ± 0.022	0.409 ± 0.074	0.568 ± 0.073
SVM	0.595 ± 0.013	1.000 ± 0.000	0.190 ± 0.026	0.319 ± 0.037

Performance

• XGBoost: Accuracy ~96%, Precision ~92%, AUC ~0.76
• Model validated using 5‑fold cross‑validation

Performance Metrics

Metric	Value	Notes
Balanced Accuracy	0.881 ± 0.017	Based on cross-validation
Precision	0.963 ± 0.007	On unseen data
Recall	0.763 ± 0.035	On unseen data
F1 Score	0.851 ± 0.023	Overall performance measure

Here's your data structure diagram:

erDiagram
    TRANSACTION ||--|| CUSTOMER : "belongs to"
    TRANSACTION ||--|| TERMINAL : "processed by"
    
    CUSTOMER {
        string CUSTOMER_ID
    }
    
    TERMINAL {
        string TERMINAL_ID
    }
    
    TRANSACTION {
        string TRANSACTION_ID
        float TX_AMOUNT
        int TX_TIME_SECONDS
        int TX_TIME_DAYS
        boolean TX_FRAUD
        string TX_FRAUD_SCENARIO
        int TX_HOUR
        int TX_DAY_OF_WEEK
        int TX_WEEK_OF_YEAR
        boolean TX_IS_WEEKEND
    }

Loading

The diagram shows how your transaction data is organized:

• Each transaction belongs to exactly one customer and is processed by exactly one terminal
• Fields are grouped into logical categories:
  • CUSTOMER: Customer identification
  • TERMINAL: Terminal identification
  • TRANSACTION: Core transaction details and fraud information

Fraud Transaction Detection
Simulated dataset (~1.8M transactions, ~0.8% fraud) generated using real‑world distribution patterns.

Here are your data attributes in table format:

Field Name	Purpose	Format	Use Case	Importance
TRANSACTION_ID	Unique identifier for each transaction	String	Transaction tracking and audit trails	Essential for fraud investigation
CUSTOMER_ID	Identifies customer account	String	Customer behavior analysis	Critical for pattern recognition
TERMINAL_ID	Identifies processing terminal	String	Terminal security monitoring	Key for location tracking
TX_AMOUNT	Transaction amount in currency	Float	Fraud detection algorithms	Primary fraud indicator
TX_TIME_SECONDS	Precise transaction timestamp	Integer	Micro-level pattern analysis	Enables rapid fraud detection
TX_TIME_DAYS	Transaction timestamp in days	Integer	Long-term pattern analysis	Supports trend analysis
TX_HOUR	Hour of day (0-23)	Integer	Peak fraud hour identification	Time-based risk assessment
TX_DAY_OF_WEEK	Day of week (1-7)	Integer	Weekly pattern analysis	Behavioral pattern recognition
TX_WEEK_OF_YEAR	Week number in year (1-52)	Integer	Seasonal pattern analysis	Long-term trend analysis
TX_IS_WEEKEND	Weekend indicator	Boolean	Weekend vs. weekday analysis	Risk profile adjustment
TX_FRAUD	Fraud status indicator	Boolean	Fraud detection training	Core target variable
TX_FRAUD_SCENARIO	Fraud type categorization	String	Fraud pattern analysis	Strategy development

Requirements

pip install -r requirements.txt

Here's your Streamlit project architecture diagram:

flowchart TD
    subgraph Client["Client Side"]
        Browser["Web Browser"]
    end
    
    subgraph Streamlit["Streamlit Application"]
        App["Streamlit App"]
        Model["Fraud Detection Model"]
        Cache["Cached Data"]
    end
    
    subgraph Storage["Data Storage"]
        Pickle["Pickle Files"]
        Data["Transaction Data"]
    end
    
    Browser -->|"HTTP Request"| App
    App -->|"Load Model"| Model
    Model -->|"Request Data"| Cache
    Cache -->|"If Not Found"| Pickle
    Pickle -->|"Load Data"| Data
    Data -->|"Cache"| Cache
    Model -->|"Predictions"| App
    App -->|"Results"| Browser
    
    classDef client fill:#f9f9f9,stroke:#333,stroke-width:2px,color:#000000
    classDef app fill:#e1f5fe,stroke:#333,stroke-width:2px,color:#000000
    classDef storage fill:#fff3e0,stroke:#333,stroke-width:2px,color:#000000
    
    class Browser client
    class App,Model,Cache app
    class Pickle,Data storage

Loading

Contributing

Fork the repositoryCreate a feature branchRun tests: pytest src/tests/Submit pull request with documentation updates##

feel free to connect. https://linkedin.com/in/siddhi-rawool-783059248