PROJECT SUMMARY

NAME

Customer Churn Prediction - Machine Learning Project

CONTEXT

In the competitive telecom industry, retaining customers is crucial for maintaining revenue and growth. Customer churn, where customers discontinue their service, poses a significant challenge. By predicting churn, companies can take proactive measures to retain customers and improve overall satisfaction.

GOAL

Develop a machine learning model to predict whether a customer will churn, enabling the company to implement targeted retention strategies.

DATA

1 CSV file (7043, 21): Customer-Churn.csv

TECHNIQUES AND LIBRARIES USED:

• Data Acquisition: CSV import with Pandas
• Data Analytics: Exploratory data analysis to understand patterns and relationships
• Data Visualization: Seaborn, Matplotlib and Plotply
• Data Preprocessing: Cleaning, encoding, and scaling with Scikit Learn, Numpy and Scipy
• Data Engineering: Feature creation, selection, and imputation
• Modeling: Logistic Regression, SVC, Random Forest, and ensemble methods (voting, stacking, boosting) with Imbalanced-Learn and Scikit Learn.
• Evaluation: Recall, precision, F1-score, cross-validation, confusion matrix, and learning curve
• Tuning: GridSearchCV/RandomizedSearchCV for parameter optimization
• Deployment: Model saving with joblib

METHODOLOGY

Data Acquisition and Preparation:

• Import Data: Loaded dataset using Pandas.
• Data Cleaning: Handled missing values, corrected data types, and dropped unnecessary columns.

Exploratory Data Analysis (EDA):

• Descriptive Statistics: Calculated summary statistics.
• Visualization: Used seaborn and matplotlib for histograms, bar plots, and correlation heatmaps.
• Target Variable Analysis: Examined churn distribution and identified class imbalance.

Data Preprocessing:

• Encoding Categorical Variables: One-hot and target encoding.
• Scaling: Standardization/normalization of continuous variables.

Modeling:

• Model Selection: Choose several classification models suitable for the task (Handling Imbalance).
• Initial Training: Baseline performance with default parameters.
• Evaluation Metrics: Prioritized recall, also considered precision, F1-score, and accuracy.
• Parameter Tuning: Grid search and random search.
• Cross-Validation: Ensured model generalizability and prevented overfitting.

Feature Engineering

• Feature Selection: Evaluated feature importance.
• Feature Creation/Deletion: Simplified the model with new and selected features.

Model Deployment:

• Final Model Selection: Based on recall and overall performance and tradeoff.
• Saving the Model: Using joblib.
• Wrapper Class: Facilitated easy deployment and integration.

Model & Business Recommendations:

• Insights and Strategies: Derived actionable insights and proposed strategies to mitigate churn and improve model accuracy.
• Implementation Plan: A plan for targeted interventions was suggested.

DATA ANALYSIS

Observations

• Churning Rate: 26.5%
• Imbalance in Target Class: 26.5% churn vs 73.5% non-churn.
• Variable Types: 17 categoricals, 3 numericals.
• Influential Factors:
  • Demographics: 'SeniorCitizen', 'Partner', 'Dependents'
  • Services: 'MultipleLines', 'InternetService', 'OnlineSecurity', 'OnlineBackup', 'DeviceProtection', 'TechSupport'
  • Account: 'Contract', 'PaperlessBilling', 'PaymentMethod', 'MonthlyCharges', 'TotalCharges', 'tenure'

Variable Relationships with Churn:

Categorical variables:

• Demographic Details:
  • 'SeniorCitizen', 'Partner', 'Dependents': Influence churn rates (e.g., seniors, singles, and independents have higher churn rates).
• Services:
  • 'InternetService': High influence for Fiber optic with a churn rate of 69.4%.
  • Optional Internet Services: decrease churn among internet customers.
• Account Details:
  • 'Contract': Month-to-month contracts see the highest churn rate (88.6%).
  • 'PaperlessBilling': Increases churn (74.9%).
  • 'PaymentMethod': Electronic check payments have the highest churn rate (57.3%).

Continuous Variables:

• 'Total Charges': Lower charges, higher churn (new customers).
• 'Tenure': Shorter tenure, higher churn. 75% of churn occurs before 30 months(new customers).
• 'Monthly Charges': Higher charges, higher churn. Notable increase in churn probability around $65/month.

Correlation Analysis:

• Internet Services: Strong correlation with optional internet services (+0.61).
• Phone Services: Strong correlation with MultipleLines and PhoneService (+0.61).

CONCLUSION

Possible Solutions to Improve Model Performance

• More Quality Data: Collect additional customer interaction data.
• Feature Engineering: Create new features.
• Data Augmentation: Generate synthetic data using SMOTE.
• Ensemble Methods: Combine predictions from multiple models.
• Dimensionality Reduction: Apply Principal Component Analysis (PCA).
• Regularization: Use L1/L2 regularization.
• Temporal Analysis: Incorporate time-based features.

Possible Business Solutions to Improve Customer Churn

Proactive Customer Management

• Procedures and warning for high-risk clients
• Early intervention programs

Incentives and Rewards

• New customer incentives
• High-paying customer incentives
• Quick surveys with incentives
• Loyalty programs

Customer Engagement and Support

• Engage unhappy customers
• Enhanced customer support
• Personalized communication
• Customer feedback loop

Service and Product Improvement

• Service Quality Improvement
• Customer Education

Data-Driven Strategies

• Predictive Analytics

Flexibility and Customization

• Flexible Contract Options
• Tailored Offers and Promotions