Churn prediction model deployment

Project aim: Creating a churn prediction model (binary classification) and deploying it using docker container on local Linux Machine and AWS Elastic Beanstalk using awsebcli.

Data Source: https://www.kaggle.com/datasets/blastchar/telco-customer-churn

Project Requirements: python 3.9, pandas, numpy, scikit-learn and docker

Operating System: Linux

Tools: Visual Studio Code and Linux command line

The project involved the following steps:

Loading dataset & EDA:

• Downloading, loading, cleaning and preparing data for analysis
• Performing exploratory data analysis
• Performing a feature importance analysis using the concept of Risk Ratio
• Performing correlation analysis for feature selection

Feature engineering:

• Performing feature engineering which involves One-Hot Encoding for categorical variables

Model training:

• Fitting and training a logistic regession binary classification model using scikit-learn's LogisticRegression class

Model Evaluation:

• Model evaluation using basic evaluation matric such as accuracy
• Identifying cut-off for achieving maximum accuracy
• Computing auc-roc (due to target class imbalance)
• Used pickle library to save and load the model

Creating a flask-based web application:

• Created a basic flask app churn-serving.py for predicting churn based on single observation input
• Tested flask app's prediction using request python library
• Used pipenv to manage project dependencies. This will create Pipenv and Pipenv.lock files. It manages exact depndencies that we need to run the application.
• Installed gunicorn using pipenv which is a production ready WSGI server tool. It can handle http request without any issues.

Dockerizing the application:

• Installed docker for creating and running a docker container
• Created a Dockerfile and added instructions
• Built the churn-prediction docker container

Making predictions on local Linux machine using request library

• Run the docker container and tests to check for model prediction

Deploying model using AWS Elastic Beanstalk

• Locally tested and deployed on AWS using Elastic Beanstalk CLI (awsebcli)
• Tested the prediction using the AWS Application host's IP address