Customer Churn Prediction

The communication technology industry is one of the most competitive industries nowadays. The main problem that practically all telecommunications industries worldwide are currently facing is customer churn.Churn, in the context of telecommunications, is the process by which customers leave a business and stop using the services it provides either because they are unhappy with those services or because they can find better options from other network providers at more reasonable prices. This could result in a loss of revenue or profit for the business. Additionally, keeping clients has grown to be a challenging task.

In order to provide their clients with the best services possible and keep them satisfied, businesses are working hard to introduce new cutting-edge applications and technology.Since losing them would result in a large loss of revenue for the business, it is imperative to identify those clients who are likely to quit the organization in the near future in advance. Accurately predicing churn can lead to higher customer retention rates, increased market share and improved business performance.This project seeks to accomplish this procedure, which is known as churn prediction.

Business is becoming extremely saturated in this competitive world. SyriaTel telecommunications company wants to take the required actions to retain customers in order to stabilize their market value because the cost of recruiting new customers is significantly higher than the cost of maintaining existing customers and also in order to reduce losses incurred from customer churn.

To predict customer churn using a classification algorithm model

To do exploratory data analysis on the data

To fit different classification algorithm models to determine which one works best for churn prediction

To select the best model

To make predictions using the selected model

To check the accuracy of the predicted variables

This data is from syriatel telecommunication company and it was obtained from kaggle Churn in Telecom's dataset.

The data has 21 columns and 3333 rows. The target column, churn, is a bool column where True means the customer did churn and False means the customer did not churn, making this a binary classification problem. The other columns were; 4 categorical and 16 numerical columns.

We started by analyzing each of the predictor variables individually by looking at their characteristics and distribution. As part of preparing our data for analysis, we cleaned it as outlined stepwise in the notebook. Univariate analysis was performed to understand the distribution of our individual variables, and later compared to our target variable, churn, in bivariate analysis so that we can understand how the target variable is distributed among the predictor variables.

We checked how the variables were related to each other and dropped the variables that were highly correlated so as to prevent multicollinearity when modeling. The categorical variables were then transformed to be numerical for modeling purposes.The data was split into training and testing sets so that we can train the model using the training set and assess it using the testing set. Different classification algorithms were fit to the data and since most of them require data to be scaled, we standardized the predictor variables in the training set and the testing set.

The data had class imbalance problems so we resolved that using SMOTE and used the resampled data to train the model. The classification models used were; logistic regression, K- nearest neighbors , decision trees and random forest. For all the models, a baseline vanilla model was fit and the score( accuracy) of the model checked, then we used gridsearch to tune the model by checking how different hyperparameters combinations affected the model’s accuracy and the best model was fit to the training data and assessed using both the training and testing data. The final model was the one that had the highest accuracy ie the decision tree classifier. It was used to make predictions and the model was evaluated using different metrics.

Graph of individual variables' distribution: The numerical predictor variables are distributed as shown:

The categorical predictor variables were distributed as shown:

Distribution of our target variable : The data had class imbalance problem ie most customers in this data did not churn.

Features importance graph: A features importance graph was drawn to determine the most influential features in the decision tree model, allowing you to understand which features have the greatest impact on the model's predictions

confusion matrix The confusion matrix of the model:

Due to the presence of high multicollinearity among several predictor variables, we had to remove some of those columns from our analysis.

The model has some limitations, since decision trees can be biased towards the majority class if the dataset used for modeling has imbalanced class distribution, we had to use SMOTE to solve the class imbalance problems thus any new data fed to the model has to undergo the same preprocessing technique.

The predictor variables;account length, total day minutes,total day calls, total eve minutes, total eve calls, total day charge, total eve charge, total night minutes, total night calls, total night charge , total intl minutes and total intl charge seem to be fairly normally distributed, total intl calls and customer service calls seem to be normally distributed but skewed to the right, area code is a discrete categorical column and number of voicemail messages, most of them were 0 but the rest seem to have a platykurtic distribution.

Wv state seems to be having most of the customers and CA has the least customers, most of the customers don't have an international plan and also most of them don't have a voicemail plan, explaining why the majority of the data in the number of voicemail messages is 0.

The most influential predictors were; customer service calls, total day minutes, International plan, voicemail plan, total eve charge, total intl charge, total intl calls,total night minutes, total eve calls, state LA, state MS, state NJ, state ME,state IL, state HI , state AL and account length(They are listed according to their impact in descending order

The model has a good balance between precision and recall. It means that the model is effectively identifying the positive class instances while minimizing false positives and false negatives. The model has 92.6% accuracy thus the model is suitable for churn prediction

The number of customer service calls was identified as one of the most influential predictors of churn. This suggests that providing excellent customer service and addressing customer concerns promptly and effectively can significantly reduce churn. Focus on training customer service representatives to handle customer issues efficiently and provide proactive support to enhance customer satisfaction and loyalty.

Total day minutes, total eve charge, total night minutes, and total international calls and charges were identified as influential predictors. Analyze these usage patterns further to identify any specific trends or behaviors associated with churn. For example, if customers who have high evening charges are more likely to churn, consider targeted offers or discounts to encourage retention during evening hours.

Most customers do not have an international plan or voicemail plan. The company should consider promoting these services to customers who currently don't have them but may benefit from them.

Since the dataset shows a significant concentration of customers in WV state and fewer customers in CA, the company should consider focusing retention efforts and marketing campaigns on these regions

The company should Leverage the influential predictors and usage patterns identified by the model to develop personalized retention campaigns. For example, for customers with a high number of customer service calls, consider reaching out to them proactively with special offers or personalized assistance to address their concerns and improve their experience.

While the model currently has good accuracy and performance, we should continue monitoring and evaluating its performance on new data. As customer behaviors and preferences change over time, it's important to ensure that the model remains effective and up-to-date.

Exploring advanced techniques like ensemble methods, gradient boosting methods, XGboost and Adaboost , or deep learning to further improve churn prediction performance