/churn_prediction

Primary LanguageJupyter Notebook

Churn Prediction Project

Summary

1. Introduction

This repository contains the first iteration of CRISP-DS to analyze and create a Churn Classification Model to help a bank classify their customers based on their data.

Predicting customer churn is extremely important to any business, because the majority of the companies with high revenue are the companies which have high retention rates. Sean Ellis and Morgan Brown indicate in their book, Hacking Growth, that when a enterprise retain their clients, they just don't share and don't loose market space for their competitor, but will also increase their revenue. The books says that with only 5% retaintion rate, a company can increase up to 25% to 95% of their profit.

This Readme goal is to show how the project construction phases and analysis were done, and also show how to use the project in order to classify new clients if they will churn or not.

2. The Dataset

The dataset of this project can be found in the data/training directory

It has 14 columns and 10.000 lines.

2.1. Data Description

Column Description
RowNumber Row ID
CustomerId Customer's ID
Surname Customer's Last name
CreditScore Customer's Credit Score
Geography Customer's Country of Residence
Gender Costumer's Gender
Age Costumer's Age
Tenure Quantity of years the customer is a bank's client client
Balande How much money the client have on it's account
NumOfProducts Quantity of products the client had purchased with the institution
HasCrCard Indicates if the client has a credit card or not with the institution. 1 - client has credit card; 0 - client dosen't have credit card
IsActiveMember Indicates if the client is active or not. 1 - active client; 0 - non active client
EstimatedSalary Client's estimated anual salary
Exited Indicates whether the client has an account at the institution. 1 - Client has exited the bank; 0 - client dosen't have exited the bank

2.2. Dataset Summary

2.2.1. Numerical Columns

attributes min max range mean median std skew kurtosis
credit_score 350 850 500 650.753 652 96.5904 -0.0795404 -0.418624
age 18 92 74 38.9466 37 10.5262 1.03571 1.45945
tenure 0 10 10 5.01663 5 2.89661 0.0131241 -1.17284
balance 0 238388 238388 76381.2 97055.1 62298.2 -0.141721 -1.49221
num_of_products 1 4 3 1.5295 1 0.581704 0.755984 0.636943
has_cr_card 0 1 1 0.70875 1 0.454338 -0.919091 -1.15556
is_active_member 0 1 1 0.514875 1 0.499779 -0.0595375 -1.99695
estimated_salary 11.58 199992 199981 99730.8 99446.9 57331.5 0.00712627 -1.17479

2.2.2. Categorical Columns

s

s

2.2.3. Terget Column

s

2.2.4. Data Summary

The actual Churn rate from the bank is 20.375%

Besides, we have the following analysis:

  • Balance: About 25% of the observations have value equal to zero. Must be investigated to validate the possible outliers.
  • credit_score feature has a spike around 850 points, with approximately 300 customers. Must be investigated to determine if these observations aren't outliers.
  • age feature indicates that the bank customers are concentrated between 30 and 45 years
  • tenure feature has around 300 customers with zero years. Theses observations must be investigated to better understand these customers
  • has_cr_card feature indicates that the majority of the bank customers have a credit card
  • is_active_member feature indicates the quantity of customers that are active and the customers that are not active are approximately the same
  • estimated_salary: The graph shows that the distribution is partially uniform. It needs to be investigated.
  • Most customers are men, but the customers who most left the bank the most are women
  • Most customers are from France, but the customers who left the bank the most are from Germany

credit_score, tenure, estimated_salaray and balance will be investigated deeply in the next iteration.

3. Feature Engineering

3.1. Hypothesis Mind Map

Was created a Mind Map to help not only on the feature creation, but also to help create Business Hypothesis in order to guide EDA process.

3.2. Feature Engineering

Were created 4 features:

  • credit_score_category: feature created to categorize customers based on their credit score
  • estimated_monthly_salary: feature created to calculate the customer's monthly income
  • products_used_year: feature created to calculate how many products the customer use by year

3.3. Feature Filtering

Were filtered row_number, costumer_id and surname features, because they don't add any information relevance to the machine learning model.

4. EDA - Exploratory Data Analysis

4.1. Bivariate Analysis - Chosen Business Hypothesis

Hypothesis Conclusion Relevance
H1. Customers with 4 years or less as bank customers are the largest number of active customers. FALSE Low
H3. Customers that purchased 2 or more products from the bank are the majority of active customers. TRUE Low
H5. Customers with Credit Score above 650 points are the majority of customers that left the bank. FALSE High
H7. Customers who have a credit card are the majority of customers who have left the bank. TRUE High
H8. Male customers who have a credit card purchased more products than female customers who have a credit card. TRUE Low

4.2. Multivariate Analysis

4.2.1. Numerical Attributes

4.2.2. Categorical Attributes

4.3. EDA Summary

  • There are no significant linear correlation between the numerical features
  • There are no significant correlation between the categorical feature.
  • credit_score has a lot of significancy

5. Data Preparation and Feature Selection

5.1. Data Preparation

For the numerical features, in this iteration, was used the Standard Scaler to rescale all features. In the next iteration, the credit_score, tenure, estimated_salaray and balance features will be investigated and, if outliers will be found, the Scaler could be changed.

For the categorical features, were used One Hot Encoding, to encode the geography and gender features because of their low cardinality.

5.2. Feature Selection

row_number, costumer_id and surname were already manualy filtered because of their low significance. In this iteration were used all features available. Besides, it was used Random Forest algorithm to help visualize feature importance. In the next iterations, it could be used another technique to select only the significant features, like RFE, PCA or Sequential Feature Selector.

6. Machine Learning Model and Fine Tunning

6.1. Finding the Best Model and Hyperparameters

In order to find the best machine learning model and its hyperparameters, the Grid Search CV object was used with 4 Machine Learning Models:

  • Logistic Regression
  • Support Vector Classification (SVC)
  • Random Forest
  • XGBoost

Besides, it was created a Pipeline to process all Machine Learning Project stages, like transformations and encodings, in order to find the best model and hyperparameters. Also, the chosen metric was roc_auc_score.

The best machine learning model found was XGBoost with 0.86 roc_auc_score.

6.2. Training and Evaluating Chosen Model

It was created a Pipeline to make all preprocessing and training with the selected Machine Learning Model.

The trained model had a roc_auc_score of 0.884 with the training data and had a roc_auc_score of 0.861with the test data.

6.3. Saving Model

To create the final model, it was created a new Pipeline that used all transformations and preprocessig pipelines to train a new model with all available data to train the chosen model, in order to help the model perform better with unseen data.

This new model was saved in the models directory with the pickle format.

6.4. Project availability preparation

In order to facilitate the project usage, it was created a Python script that will load the model saved on pickle, read the test csv file, execute the machine learning model with the new unseen test data and create a response csv file, with the unseen data classified.

7. How to Use this Project

In order to use this project, you need to execute these steps:

  1. Create a Python virtual environment with Python's version 3.10
  2. Install all packages and libraries inside requirements.txt using the pip install -r requirements.txt command
  3. Replace the CSV file that is inside data/test/ directory with the file that have the data you want to classify. The CSV file MUST have exactly the abandono_teste.csv name.
  4. The CSV file with the classified clients will be placed inside the data/answer/ directory, under the name abandonos_{date}.csv, where {date} will be the date of the execution.

8. Next Steps

In the next iterations, credit_score, tenure, estimated_salaray and balance features will be deeply investigated to better understand the data and verify the existence of outliers that could be filtered.

Another thing that can be done is the creation of new features to enrich the data.

Besides, we can use feature selection techniques to choose the most relevant features, in order to increase the model score.

We can also answer the other business hypothesis to better understand the data and the features realations and create new hypothesis the verify other features relations.

Back to top