Skip to content

Latest commit

 

History

History
124 lines (98 loc) · 4.41 KB

English_README.md

File metadata and controls

124 lines (98 loc) · 4.41 KB

Netflix Churn Analysis

  • This project builds an end-to-end machine learning pipeline to predict customer churn for a subscription-based digital service.
  • The goal is to identify at-risk users early using behavioral, engagement, and subscription features.
※ Report Findings are summarized in the business_report section of the repository

Problem Statement

Customer churn is a critical business problem for subscription-based services. This project aims to identify users likely to churn by analyzing behavioral, engagement, and subscription-related data.

The goal is not only prediction accuracy, but also interpretability and business-driven insights that can support retention strategies.

Key objectives:

  • Analyze user behavior patterns linked to churn
  • Engineer meaningful engagement features
  • Train and evaluate regression model
  • Interpret model outputs for business analytics report

Dataset Overview:

Netflix-Style Synthetic Dataset

  • Dataset Overview - This synthetic dataset simulates a Netflix-style streaming platform with realistic user behavior, content catalog, and engagement metrics. Perfect for machine learning, data science, and analytics projects.

  • Key Features of "users.csv" - 10,300 total records with User demographics, subscription plans, regional data

  • Key Features of "watch_history.csv" - 105,000 total records with viewing sessions with device, quality, progress data

  • Key Features

    • The Final pruned dataset consists of:
    • Size: 8,431 users
    • 17 engineered features
    • Target Variable:
      • churn (1 = churned, 0 = retained)

Feature Groups

Category Features
Demographics age, gender, country
Subscription subscription_plan, monthly_spend
Engagement avg_session_minutes, num_sessions, watch_per_day
Retention Signals tenure_days, days_since_last_watch
Content Completion completion_rate
Engineered Buckets watch_bucket, inactivity_bucket, completion_bucket
Value Metrics value_ratio

Feature Engineering

Key engineered features include:

  • Engagement intensity

    • watch_per_day
    • avg_session_minutes

  • Inactivity modeling

    • days_since_last_watch
    • inactivity_bucket

  • Consumption quality

    • completion_rate
    • completion_bucket

  • Customer value

    • value_ratio (watch time relative to spend)

These features are designed to capture early disengagement patterns rather than relying solely on static demographic attributes.

Modeling Approach

  • Model: Logistic Regression
  • Reason:
    • Strong baseline for churn problems
    • Interpretable coefficients
    • Easy communication with business stakeholders

Model Evaluation

Classification Performance

Metric Value
Accuracy 0.50
ROC-AUC 0.49
Recall (Churn) 0.49
Precision (Churn) 0.14

The model captures nearly half of churned users (high recall) but produces many false positives, indicating weak linear separability in the data.

Visual Diagnostics

ROC Curve

Confusion Matrix

Feature Coefficients

Business Interpretation

  • Suitable for early churn warning systems
  • High recall supports proactive retention campaigns
  • False positives are acceptable in marketing-driven interventions
  • Indicates need for nonlinear or temporal models for better discrimination

Limitations & Future Work

  • Logistic regression struggles with nonlinear behavior patterns
  • Class imbalance affects precision
  • Temporal behavior trends are not explicitly modeled

Future improvements:

  • Gradient boosting models (XGBoost, LightGBM)
  • Cost-sensitive learning
  • Time-series engagement features
  • SHAP-based explainability

Frameworks

  • Python,
  • Pandas,
  • Scikit-learn,
  • Matplotlib,
  • Seaborn,
  • Jupyter Notebook