Maven Music Churn Analysis

Maven Music: Customer Churn Analysis and Data Preparation

Maven Music, a music streaming service, has been experiencing higher-than-usual customer churn. This project applies the complete data science workflow—from data gathering through feature engineering—to investigate churn patterns and prepare a modeling-ready dataset using three months of customer subscription and listening history data.

The complete project repository and notebook can be found at Maven Music Churn Analysis Repository

Feature Engineering for Predictive Modeling

The feature engineering process created a modeling-ready dataset with one row per customer and the following features:

Target Variable

• Cancelled: Binary indicator (1 = cancelled, 0 = active)

Engineered Features

1. Discount? (Binary, 0/1)

   Indicates whether customer received a subscription discount

2. Num Sessions (Integer)

   Count of listening sessions per customer, calculated using:

   num_sessions = history.groupby('Customer ID')['Session ID'].count()

   Range: 2-60 sessions across customers

3. Percent Pop (Float, 0-100)

   Percentage of listening history consisting of Pop music

   Calculation steps:

   • Created dummy variables for each genre using pd.get_dummies()
   • Grouped by Customer ID and summed genre counts
   • Calculated total audio items per customer
   • Computed percentage: (Pop count / Total Audio) * 100
4. Percent Podcasts (Float, 0-100)

   Percentage of listening history consisting of podcasts (Comedy and True Crime genres)

   Calculated as: ((Comedy + True Crime) / Total Audio) * 100

Data Integration

The final modeling dataset was constructed through systematic merging:

1. Started with customer base features (Customer ID, Cancelled, Discount?)
2. Merged listening session counts
3. Integrated genre preference percentages
4. Result: Clean DataFrame with all numeric, non-null features ready for machine learning

Correlation Analysis and Key Findings

Correlation analysis of the modeling features revealed which variables have the strongest relationship with customer churn:

Feature Correlations with Churn

Key Insights

• Most Predictive Features: Discount? and Percent Pop show the highest correlation with churn, making them the most valuable features for predictive modeling
• Genre Preferences: Customers with higher Pop listening percentages are significantly more likely to cancel, suggesting potential content or recommendation issues for this segment
• Engagement Patterns: Lower session counts correlate with higher churn, though the relationship is relatively weak
• Discount Paradox: Customers receiving discounts show higher churn rates, potentially indicating these were retention attempts for already at-risk customers

Additional Feature Relationships

• Percent Pop and Percent Podcasts have strong negative correlation (-0.487), indicating customers tend to specialize in one content type
• Num Sessions shows weak correlation with Percent Pop (0.210), suggesting engagement level has minimal relationship with genre preference
• Num Sessions and Percent Podcasts are negatively correlated (-0.495), suggesting podcast listeners may have shorter but more focused sessions

Next Steps and Extensions

The modeling-ready dataset enables several natural extensions:

• Building classification models (logistic regression, decision trees, random forests) to predict churn probability
• Developing customer segmentation using clustering on listening behavior patterns
• Creating time-based features from session timestamps to capture temporal engagement patterns
• Expanding feature set with additional behavioral metrics (session duration, skip rates, playlist usage)
• A/B testing retention interventions targeted at high-risk segments

This analytical framework applies broadly to subscription-based businesses facing similar churn challenges across streaming media, SaaS platforms, and digital content services.

Conclusion

This project demonstrates a complete data preparation and EDA workflow for customer churn analysis. Starting with raw CSV and Excel data, the analysis systematically cleaned data, handled missing values, engineered behavioral features, and identified key predictors of churn.

The correlation analysis revealed that Discount? (r=0.47) and Percent Pop (r=0.59) are the strongest predictors of cancellation, providing clear direction for future predictive modeling efforts. The modeling-ready dataset—with 30 customer records and 5 numeric, non-null features—is prepared for classification algorithms such as logistic regression, decision trees, or random forests.

Beyond preparing data for modeling, the analysis surfaced actionable insights: customers who listen primarily to Pop music and those receiving discounts show significantly higher churn rates. These findings suggest targeted retention strategies could focus on improving content recommendations for Pop listeners and re-evaluating discount effectiveness.

This work exemplifies foundational data science skills: data gathering from multiple sources, systematic data cleaning, exploratory analysis with visualization, feature engineering, and preparation of modeling-ready datasets—all documented in a reproducible Jupyter notebook workflow.

Project Context and Attribution

This project serves as the final project for Maven Analytics' Data Science in Python: Data Prep & EDA course, demonstrating comprehensive skills in data gathering, cleaning, exploratory analysis, and feature engineering.

Course Information

• Course: Data Science in Python: Data Prep & EDA
• Platform: Maven Analytics
• Instructor: Alice Zhao, Data Scientist and Author of SQL Pocket Guide (O'Reilly)
• Topics Covered: Project scoping, data gathering, data cleaning, EDA, feature engineering, and data preparation for modeling