Adaptive Fraud Intelligence

Decision-centric fraud detection system with adaptive alert selection, cost-aware ranking, operational monitoring, and real-time simulation support.

---

Overview

Traditional fraud detection systems often optimize only predictive performance metrics such as ROC-AUC or precision.
This project focuses instead on operational fraud decision-making under real-world investigation constraints.

The system combines machine learning fraud scoring with adaptive decision logic, investigation-budget management, analyst prioritization, and cost-aware alert selection to simulate realistic fraud operations workflows.

---

Core Objectives

• Detect fraudulent financial transactions
• Reduce total operational fraud cost
• Improve fraud recall under investigation constraints
• Prioritize analyst review queues intelligently
• Simulate adaptive fraud monitoring workflows
• Compare static threshold systems against adaptive decision systems

---

Features

Machine Learning Layer

• Fraud probability scoring
• Transaction risk estimation
• Threshold evaluation
• Feature engineering pipeline
• Probability calibration support

Decision Intelligence Layer

• Adaptive alert budgeting
• Cost-aware fraud ranking
• Investigation-capacity-aware alert selection
• Static threshold vs adaptive strategy comparison
• Risk-zone transaction prioritization
• Analyst queue optimization

Monitoring & Simulation

• Sequential fraud simulation
• Operating curve analysis
• Monitoring metrics
• Real-time simulation support
• Alert volume analysis
• Operational performance tracking

Interfaces

• FastAPI backend
• Streamlit interactive dashboard

---

Main Result

Compared to the static threshold baseline, the adaptive decision system achieved improved fraud recall while simultaneously reducing total operational cost.

System	Recall	Precision	Alerts	Total Cost
Static Threshold	0.625	0.033	303	600,060
Decision System	0.875	0.038	372	552,563

Key Outcome

The adaptive decision engine:

• Increased fraud recall by 25%
• Reduced total operational cost
• Maintained operationally manageable alert volumes
• Improved analyst prioritization efficiency

---

Dashboard Preview

Static Threshold vs Decision System

Operating Curve Analysis

Analyst Investigation Queue

---

System Architecture

flowchart TD
    A[Raw Transaction Dataset] --> B[Data Loading & Preprocessing]
    B --> C[Fraud Scoring Model]
    C --> D[Fraud Risk Score]

    D --> E[Decision Engine]

    E --> F[Static Threshold Baseline]
    E --> G[Adaptive Decision System]

    G --> H[Risk-Zone Prioritization]
    G --> I[Cost-Aware Ranking]
    G --> J[Alert Budgeting]
    G --> K[Suppression Logic]

    H --> L[Analyst Queue]
    I --> L
    J --> L
    K --> L

    L --> M[FastAPI Backend]
    M --> N[Streamlit Dashboard]

    N --> O[Operating Curve Analysis]
    N --> P[Monitoring Metrics]
    N --> Q[Alert Review Interface]

Loading

The architecture separates the predictive layer from the decision layer.

The machine learning model produces fraud risk scores, while the decision engine converts those scores into operational actions using adaptive alert budgets, cost-aware ranking, and risk-zone prioritization.

---

Project Structure

app/
    api/
    core/
    model/
    monitoring/
    tests/

decisioning/
    cost_logic.py
    decision_engine.py
    strategies.py
    suppression.py
    thresholding.py

scripts/
    evaluate_decision_strategies.py
    run_sequential_simulation.py
    run_realtime_demo.py
    plot_operating_curve.py

config/
models/
notebooks/
docs/

---

## Technologies

- Python
- Pandas
- NumPy
- Scikit-learn
- FastAPI
- Streamlit
- Matplotlib
- Joblib

---

## How To Run

### Install dependencies

```bash
py -m pip install -r requirements.txt

Add dataset

Place dataset here:

data/raw/AIML Dataset.csv

Run FastAPI backend

py -m uvicorn app.api.main:app --reload

Run Streamlit dashboard

py -m streamlit run app/dashboard.py

---

Dataset

This project uses the PaySim synthetic financial transaction dataset.

The dataset is not included in the repository due to size limitations.

Expected dataset path:

data/raw/AIML Dataset.csv

---

Future Improvements

• Probability calibration improvements
• Drift detection
• Online learning
• Analyst feedback loops
• Queue-aware investigation allocation
• Dynamic risk adaptation
• Real streaming integration

---

Thesis Context

This repository was developed as part of a Master's thesis focused on decision-centric fraud detection, adaptive transaction prioritization, cost optimization, and monitoring for real-time capable fraud systems.