Insurance Fraud Detection

Implementation of Insurance fraud detection.

Get Started

Data set has been taken from 2018 UMN STAT 8051 Modeling Kaggle competition. Train and test sets can be found at ./data/train.csv and ./data/test.csv

To run the application locally, you require Python>3.8 and pip

Use pip to install the packages with below command:

pip install -r requirements.txt

Run the application

This application requires the following packages to start a development server:

1. fastAPI
2. Uvicorn

Run the below command to start the application:

uvicorn api:app

Note: Use --reload to reload the server on code changes

API

Currently there are two APIs that can train and infer the model. Below are the two API description:

1. /train

   train is a HTTP Post call that takes params such as depth, features and number of estimators to name a few. Pickle has been used to save the best model at ./models/store_best_model.pickle.

2. /test

   Uses the saved model and finds out if the claim is a fraud or not.

More info and API documentation can be found at:

http://localhost:8000/docs.

Zipped file contents

After training, the best model will be saved in models. The report.json file containing the results will be saved in the output folder after testing using test data.