Insurance Fraud Claims Detection

This project focuses on building a model to predict fraudulent vehicle insurance claims.

Files

• Dataset:

  • insurance_claims.csv

• Python Code:

  • Jupyter notebook: insurance_claims.ipynb

Problem Statement

The objective is to develop a model capable of predicting fraudulent vehicle insurance claims. The dataset is obtained from Kaggle: Insurance Fraud Claims Detection.

Data Analysis

• Dataset Overview:

  • Contains 1000 observations and 40 dimensions.
  • Binary classification problem - 'fraud_reported' (Yes/No).

• Column Descriptions:

  • Columns include various features such as months_as_customer, age, policy_number, etc.

• Missing Value Analysis:

  • Identified missing values represented as "?" and replaced with NaN.
  • Dropped variables with missing values exceeding 30%: _c39, property_damage, police_report_available.

• Feature Engineering:

  • Converted date columns (policy_bind_date, incident_date) to datetime format.
  • Derived new features like policy_Month, policy_Year, Inc_Month.
  • Created Capital_Loss_or_Gain based on capital-gains and capital-loss.
  • Imputed missing values for collision_type using KNN Imputation.
  • Created a new variable total_claim and converted related variables to categorical type codes.

• Feature Selection:

  • Applied correlation analysis (Heatmap) to identify and address multicollinearity.
  • Dropped months_as_customer due to high correlation with age.
  • Applied Chi-Sq-Test to select significant categorical variables.

• Down-sampling:

  • Balanced the imbalanced dataset using resampling techniques (0 - 400, 1 - 247).

• Model Building:

  • Random Forest:

    • Accuracy: 0.8307692307692308
    • Confusion Matrix:

      [[110  22]
      [11  52]]
      

  • Decision Tree:

    • Accuracy: 0.7487179487179487
    • Confusion Matrix:

      [[108  24]
      [25  38]]
      

  • Logistic Regression:

    • Accuracy: 0.6564102564102564
    • Confusion Matrix:

      [[119  13]
      [54  9]]
      

  • Random Forest performed the best among the models.

Usage

1. Clone the repository:

   git clone <repository_url>

2. Navigate to the project directory:

   cd Insurance-Fraud-Claims-Detection

3. Explore the dataset and Jupyter notebook for detailed analysis and code.

Feel free to refer to insurance_claims.ipynb for an in-depth explanation of the project.

Note: Adjust file paths and comments as needed for your project structure.