Telecom Customer Churn Prediction using Machine?Learning

Abstract

In the telecommunications industry, predicting customer churn is critical for improving retention strategies and sustaining growth. This project develops a machine learning model to accurately predict telecom customer churn using a dataset that includes demographic, service usage, and customer interaction data. Various models, including Logistic Regression, Decision Trees, Random Forests and K-Nearest Neighbors were implemented and evaluated using key metrics like accuracy, recall, and F1-score. Through rigorous data cleaning, dummy variable creation, and feature engineering, the model’s performance was optimized. The project highlights the importance of data-driven approaches to identify at-risk customers, reduce churn, and foster long-term customer relationships, ultimately promoting business growth and financial stability.

1. Introduction

Customer churn, the process where customers stop using a company’s products or services, is a significant challenge in the telecom industry. With high competition and numerous service providers offering similar features, retaining customers is often more cost-effective than acquiring new ones. Telecom companies must, therefore, leverage data analytics to identify potential churners early and take corrective actions to enhance customer satisfaction and loyalty.

This project focuses on building a predictive model to accurately forecast which customers are at risk of churning. The dataset, sourced from Kaggle, contains customer demographics, service usage patterns, and interaction history, providing a rich foundation for analysis. Using this data, we aim to extract valuable insights that can help businesses reduce churn rates.

Through the implementation of machine learning techniques such as Logistic Regression, Decision Trees, Support Vector Machines, and Random Forests, we explore various predictive models to identify the most effective approach. In this blog, we will walk through the data cleaning process, dummy variable creation, feature engineering, model building, and evaluation of results. The ultimate goal is to create a reliable tool that telecom companies can use to predict churn and foster long-term customer relationships, leading to enhanced financial sustainability.

2. Methodology

Effective data cleaning is essential for ensuring the quality and accuracy of predictive models. In this project, the dataset presented several challenges, including missing values, skewed distributions, and categorical data that required proper handling before model building. Here’s a detailed breakdown of the data cleaning process:

A pie chart was created to examine the balance of the data, revealing an imbalance where a significant number of individuals do not churn.

Handling Missing Values

• ServiceArea was the only categorical column with missing data. Missing values were imputed using the most frequent category (mode), assuming that these values represented a typical service area rather than anomalies.
• Rows with missing values in MonthlyRevenue also had missing values in related columns such as MonthlyMinutes, TotalRecurringCharge, DirectorAssistedCalls, OverageMinutes, and RoamingCalls. Additionally, PercChangeMinutes and PercChangeRevenues were missing in these rows. Patterns were explored, and missing values were either imputed based on related features or rows were removed where data was insufficient.

Correlated Missing Data

• When MonthlyRevenue was not null, only 211 values of PercChangeMinutes and PercChangeRevenues were missing. These values were imputed by examining other closely related service usage columns, ensuring consistency in the data.

Treating Categorical Variables

• The IncomeGroup variable, initially represented numerically, was treated as a categorical variable. This decision was based on its representation as income brackets rather than continuous numerical data. Furthermore, the distribution of this and other variables exhibited a strong right skew, which could negatively impact model performance. To address this, a log transformation was applied to "stretch out" the tails and normalize the distribution, making it easier for machine learning algorithms to interpret.

Outliers

• Outliers, particularly in numerical columns such as revenue and service usage, were detected using statistical methods like the Interquartile Range (IQR). Extreme outliers were either capped or removed based on their influence on the overall data distribution, ensuring that the model would not be skewed by atypical values.

By addressing missing data, transforming skewed variables, and treating outliers, a clean and reliable dataset was created, which is crucial for building accurate predictive models. The next step involved creating dummy variables for categorical columns to prepare the data for machine learning algorithms.

One Hot Encoding

To convert categorical variables into a machine learning-friendly format, we applied one-hot encoding to multiple columns. Here’s a summary of the encoded variables:

PrizmCode: Encoded categories like Other, Rural, Suburban, Town.

Occupation: Encoded categories such as Clerical, Crafts, Homemaker, Professional, Retired, Self, Student.

MaritalStatus: Encoded categories such as No, Unknown, Yes.

ServiceArea: Encoded categories like AIR, ATL, CHI, HOU, and other specific regional codes.

Homeownership: Encoded into binary values (0 and 1), indicating whether the customer owns a home.

These transformations provided a robust numerical foundation for the machine learning algorithms, ensuring that the categorical data was properly represented in the model.

Feature Engineering

Feature engineering is vital for enhancing the predictive power of machine learning models. In this project, significant steps included performing feature importance analysis, where features with scores of 0.003 or higher were retained, ensuring a focus on relevant variables while reducing noise.

In the feature engineering process, highly correlated features were identified and subsequently dropped to enhance model performance and avoid multicollinearity. Specifically, the following features were removed from the dataset:

• ActiveSubs_0.0
• Handsets_1.0
• Homeownership_1
• RespondsToMailOffers
• TotalRecurringCharge

Dropping these features was essential as their high correlation with other variables could lead to redundancy, making it difficult for the model to distinguish their individual effects on customer churn. By analyzing the correlation matrix, the decision was made to streamline the feature set, ensuring that only the most informative and independent variables remained. This refinement not only simplifies the model but also improves interpretability and predictive accuracy.

3. Result

The performance of various machine learning models was evaluated to predict customer churn, with accuracy serving as the primary metric for comparison. Below are the accuracy results obtained from each model:

• Linear Regression Accuracy: 0.7205
• Logistic Regression Accuracy: 0.7197
• Decision Tree Accuracy: 0.6066
• Random Forest Accuracy: 0.7108
• K-Nearest Neighbors (KNN) Accuracy: 0.6697

Analysis of Results

Linear Regression:

The Linear Regression model achieved an accuracy of 72.05%, making it one of the top performers in this analysis. While primarily designed for continuous outcomes, it showed reasonable performance in this classification context, suggesting that a linear relationship may exist between the features and customer churn.

Logistic Regression:

Close to Linear Regression, the Logistic Regression model attained an accuracy of 71.97%. This is a strong result, especially since Logistic Regression is specifically designed for binary classification problems. The slight drop in accuracy compared to Linear Regression may indicate some level of complexity in the data that is not entirely captured by the logistic function.

Random Forest:

The Random Forest model produced an accuracy of 71.08%. This model typically excels in handling complex datasets with high dimensionality and non-linear relationships. However, its performance here was slightly lower than expected, possibly due to overfitting or the presence of noise in the data.

K-Nearest Neighbors (KNN):

The KNN model yielded an accuracy of 66.97%, making it the least effective among the evaluated models. This result could be attributed to the model's sensitivity to the choice of the number of neighbors and the curse of dimensionality, which can hinder its performance in datasets with many features.

Decision Tree:

The Decision Tree model had the lowest accuracy at 60.66%. This result underscores a common issue with Decision Trees, which may overfit the training data or fail to generalize well to unseen data. Although Decision Trees can capture non-linear relationships effectively, their performance is heavily dependent on how the tree is constructed and the features selected.

4. Conclusion

The results highlight the effectiveness of both Linear and Logistic Regression models for predicting customer churn in the telecom industry. Despite the advantages of ensemble methods like Random Forests, their performance was not significantly better than simpler models in this case, suggesting that the data might not have sufficient complexity to warrant the added intricacies of more advanced algorithms.

The relatively lower accuracy of KNN and Decision Tree models indicates a need for further optimization, such as hyperparameter tuning and feature selection, to improve their predictive capabilities. Additionally, exploring other metrics like recall, precision, and F1-score could provide a more comprehensive view of each model's performance, particularly in a churn prediction context, where false negatives (failing to identify a churner) can be more costly than false positives.

In conclusion, the findings emphasize the importance of model selection based on the specific characteristics of the dataset and the problem at hand. The insights gained from this analysis can guide telecom companies in adopting appropriate predictive modeling techniques to enhance customer retention strategies effectively.

For more information, follow the link: https://github.com/prajwal6846/Customer-Churn-Prediction-using-Machine-Learning