Skip to main content

​
What is the Fraud Detection System?

The fraud detection system is an end-to-end machine learning application designed to identify fraudulent insurance claims. Built with Flask and powered by advanced ML algorithms, it provides automated fraud detection through a REST API interface.
The system uses a hybrid approach combining unsupervised clustering (K-Means) with supervised classification (XGBoost and SVM) to achieve high accuracy in fraud detection.

​
High-Level Workflow

The system operates through two main pipelines:

​
Training Pipeline

  1. Data Validation - Validates incoming CSV files against schema (schema_training.json)
  2. Data Preprocessing - Handles missing values, encodes categorical features, scales numerical features
  3. K-Means Clustering - Groups similar insurance claims using elbow method for optimal clusters
  4. Model Training - Trains XGBoost and SVM models on each cluster
  5. Model Selection - Selects best performing model per cluster based on AUC/accuracy scores
  6. Model Persistence - Saves trained models for prediction

​
Prediction Pipeline

  1. Data Validation - Validates prediction data against schema (schema_prediction.json)
  2. Preprocessing - Applies same transformations as training
  3. Cluster Assignment - Uses saved K-Means model to assign clusters
  4. Prediction - Applies cluster-specific best model
  5. Output Generation - Returns fraud predictions (Y/N) in CSV format

​
Key Technologies

Flask

Web framework providing REST API endpoints (/train, /predict) and monitoring dashboard integration

XGBoost

Gradient boosting classifier with hyperparameter tuning via GridSearchCV for binary fraud classification

SVM

Support Vector Machine with RBF/sigmoid kernels, used as alternative classifier in model selection

K-Means

Unsupervised clustering algorithm that segments claims into groups before classification

​
Additional Libraries

  • scikit-learn - Model training, preprocessing, and evaluation
  • Pandas & NumPy - Data manipulation and numerical operations
  • Flask-MonitoringDashboard - Performance monitoring and API analytics
  • imbalanced-learn - Handling imbalanced datasets with RandomOverSampler

​
Use Cases for Insurance Fraud Detection

Screen incoming insurance claims automatically before manual review, flagging high-risk claims for detailed investigation.
Process large batches of historical claims to identify patterns and potentially fraudulent activities that were previously missed.
Integrate predictions into risk assessment workflows to prioritize investigator resources on claims most likely to be fraudulent.
Leverage clustering approach to identify unusual claim patterns and emerging fraud schemes across different customer segments.

​
System Features

  • Schema-based Validation - Enforces data quality with JSON schema validation
  • Automated Preprocessing - Handles missing values, categorical encoding, and feature scaling
  • Cluster-based Modeling - Trains specialized models for different claim segments
  • Model Comparison - Automatically selects best performing algorithm (XGBoost vs SVM)
  • REST API Interface - Easy integration with existing systems
  • Monitoring Dashboard - Track API performance and usage metrics
  • Batch Predictions - Process multiple claims efficiently
Production Ready - The system includes comprehensive logging, error handling, and is ready for deployment with Gunicorn/Nginx.

​
Next Steps

Architecture

Explore the technical architecture and module organization

Data Pipeline

Deep dive into data ingestion and transformation processes

Fraud Detection

Learn about the fraud detection methodology and features

Build docs developers (and LLMs) love

Get started for freeTalk to us