Insurance Fraud Detection

Overview

This project focuses on detecting fraudulent insurance claims using a complete machine learning pipeline. From data preprocessing and exploratory analysis to model training and dashboard visualization, the objective was to build a reliable and explainable system for flagging suspicious claims.

Key Highlights

Cleaned and preprocessed data using LabelEncoder, outlier removal, and scaling
Trained Logistic Regression, Random Forest, and SVM models
Applied SMOTE to balance imbalanced fraud and non-fraud classes
Evaluated performance before and after SMOTE using class-wise accuracy
Visualized KPIs and insights using an interactive Power BI dashboard

BI Dashboard

The Power BI dashboard provides a visual breakdown of fraud vs. non-fraud claims, approval KPIs, and geographic risk patterns.

Before & After SMOTE: Model Evaluation

Imbalanced datasets often bias models toward the majority class. To address this, we used SMOTE (Synthetic Minority Oversampling Technique), which synthetically generates samples for the minority class (fraud cases) to improve class balance and predictive fairness.

Class-wise Accuracy Comparison

Model	SMOTE	Class 0 Accuracy	Class 1 Accuracy	Overall Accuracy
Logistic Regression	No	0.98	0.55	0.83
Random Forest	No	0.99	0.72	0.91
SVM	No	0.97	0.56	0.84
Logistic Regression	Yes	0.90	0.88	0.89
Random Forest	Yes	0.96	0.93	0.94
SVM	Yes	0.92	0.90	0.91

As shown, after applying SMOTE, fraud class (1) accuracy increased significantly across all models, especially in SVM and Random Forest, improving overall reliability for rare-event prediction.

Tools & Libraries Used