🎵 Spotify Churn Prediction System

📌 Project Overview

Customer Churn (users cancelling subscriptions) is a critical metric for streaming services. This project builds an end-to-end Machine Learning pipeline to predict which users are likely to cancel, allowing the business to intervene with retention offers.

The system handles severe Class Imbalance (80% stay / 20% leave) and utilizes a custom preprocessing pipeline with intensive Feature Engineering.

📊 Key Results Summary

Model	Accuracy	Churn Recall (Safety)	Churn Precision (Trust)	Verdict
Baseline (Guessing)	81%	0%	0%	Useless
Logistic Regression (Balanced)	68%	62%	32%	Deployed

Business Impact: While the model generates some false alarms (Precision 32%), it successfully identifies 62% of at-risk users who were previously invisible. We prioritized Recall to maximize retention opportunities.

---

📈 Exploratory Data Analysis (EDA)

Before modeling, we investigated the dataset to understand key relationships and distributions.

1. Feature Correlations

An analysis of how numerical features correlate with each other and the target variable (is_churned).

2. Numerical Feature Distributions

Visualizing the spread of key numerical behaviors like daily listening minutes and skip behavior.

3. Categorical & Geographic Analysis

Investigating how churn behaves across different regions and subscription tiers. Categorical feature breakdown:

---

🛠️ The Pipeline

1. Data Cleaning & Engineering

• String Sanitization: Fixed hidden whitespace bugs in categorical columns (e.g., ' India ' → 'India').
• Feature Engineering:
  • skips_ratio: Calculated skips_per_day / avg_daily_minutes to measure user frustration.
  • engagement_score: A weighted index of login frequency and listening time.
  • is_inactive: Flagged users with >14 days since last login.

2. Preprocessing Architecture (ColumnTransformer)

We used a split-pipeline approach to handle mixed data types safely:

• Numerical: Imputation (Median) → Standard Scaling (StandardScaler).
• Categorical: One-Hot Encoding (drop='first' to avoid multicollinearity).
• Binary Flags: Passthrough (preserved 0/1 logic without scaling).

3. Modeling

• Addressed imbalance using class_weight='balanced' in Logistic Regression.

---

📊 Model Evaluation

Confusion Matrix

The confusion matrix visualizes the performance of the deployed model on the test set. Given the imbalanced data, we focus heavily on the True Positives (Bottom Right)—the churners we successfully caught—versus the False Negatives (Bottom Left)—the churners we missed.

---

Detailed Classification Report

The detailed breakdown of Precision, Recall, and F1-Score for each class (0 = Retained, 1 = Churned).

              precision    recall  f1-score   support

           0       0.89      0.69      0.78       163
           1       0.32      0.62      0.42        37

    accuracy                           0.68       200
   macro avg       0.60      0.66      0.60       200
weighted avg       0.78      0.68      0.71       200

📂 Project Structure

Spotify-Customer-Churn/
├── EDA/                    # Exploratory Data Analysis graphs
│   ├── cat_cols.png
│   ├── correlation_matrix.png
│   ├── country_wise.png
│   └── num_cols.png
├── results/                # Model evaluation metrics/graphs
│   └── cm.png
├── data.csv                # Raw data (often gitignored in real projects)
├── X_processed.csv         # Processed features placeholder
├── main.ipynb              # Main workflow notebook
├── old_main.ipynb          # Previous versions
├── README.md               # Project Documentation
└── .gitignore              # Git exclusion rules