Spotify_Churn_Analysis

🎵 Spotify Churn Prediction – End-to-End ANN Project with Deployment

Deployed on Streamlit: View Live App

Python | TensorFlow/Keras | ANN | StandardScaler | ColumnTransformer

---

📑 Table of Contents

• Project Overview
• Dataset
• Data Exploration & Visualization
• Data Preprocessing
• Model Building & Training
• Hyperparameter Tuning
• Final Model Performance
• Predictions
• Key Takeaways
• Technologies & Libraries Used

---

Project Overview

This project predicts whether a Spotify user will churn based on their usage patterns, session behavior, and subscription metrics.
The goal is to help Spotify identify users at risk of churn and take retention actions to improve user engagement.

---

Dataset

• Source: Kaggle Spotify Dataset
• Rows: 8,000
• Columns: 12
  • user_id (int)
  • gender (object)
  • age (int)
  • country (object)
  • subscription_type (object)
  • listening_time (int)
  • songs_played_per_day (int)
  • skip_rate (float)
  • device_type (object)
  • ads_listened_per_week (int)
  • offline_listening (int)
  • is_churned (target, int: 0 = No, 1 = Yes)

---

Data Exploration & Visualization

• Checked for missing/null values → none found.
• Explored numerical distributions with boxplots.
• Explored categorical features with countplots.
• Visualized target variable with a pie chart to check class balance.

---

Data Preprocessing

• Dropped user_id as irrelevant.
• Split dataset into train (75%) and test (25%) sets.
• Used StandardScaler for numerical features and OneHotEncoder for categorical features.
• Implemented a ColumnTransformer to combine scaling and encoding.
• Saved preprocessing pipeline as processing.pkl for deployment consistency.

---

Model Building & Training

• Built an Artificial Neural Network (ANN) using Keras:
  • Input layer: 128 neurons, activation = 'elu'
  • Hidden layers: 64, 32 neurons, activation = 'elu'
  • Output layer: 1 neuron, activation = 'sigmoid'
• Compiled with binary_crossentropy loss and Adam optimizer.
• Trained for 100 epochs with EarlyStopping (patience=20) monitoring validation loss.
• Achieved validation accuracy ~0.748.

---

Hyperparameter Tuning

• Wrapped ANN in KerasClassifier for use with GridSearchCV.
• Tuned:
  • Number of neurons in each hidden layer
  • Activation functions (relu, elu)
  • Optimizers (adam, rmsprop)
  • Learning rate, batch size, epochs

Best parameters found:

{
    'n_hidden1': 64,
    'n_hidden2': 32,
    'n_hidden3': 16,
    'activation': 'elu',
    'optimizer': 'rmsprop',
    'learning_rate': 0.001,
    'batch_size': 64,
    'epochs': 20
}

---

Final Model Performance

Metric	ANN Model
Accuracy	0.748
Precision	0.75
Recall	0.74
F1-Score	0.745

• The trained ANN outperforms simple models due to capturing complex user behavior patterns.

---

Predictions

• Users input features such as:

  • Gender, Age, Country
  • Subscription Type
  • Listening Time, Songs Played per Day
  • Skip Rate, Device Type
  • Ads Listened, Offline Listening

• The trained ANN predicts the likelihood of churn and outputs a probability.

---

Key Takeaways

• ANN effectively captures non-linear relationships in user behavior.
• Preprocessing with scaling and encoding ensures consistency in deployment.
• Hyperparameter tuning significantly improved model performance.
• Real-time predictions via Streamlit deployment allow business teams to identify at-risk users.

---

Technologies & Liraries Used

• Python 3.10
• pandas, numpy
• scikit-learn (StandardScaler, OneHotEncoder, ColumnTransformer, train_test_split)
• tensorflow / keras (Sequential, Dense, EarlyStopping)
• scikeras.wrappers.KerasClassifier, GridSearchCV
• matplotlib, seaborn
• pickle
• streamlit