Project Overview:-

Project: Spotify User Subscription Prediction

Introduction

This project aims to predict user subscription status (Active/Inactive) for Spotify using a realistic dataset. The goal is to develop a robust classification model that can identify users at risk of churning or those likely to remain active subscribers. This predictive capability can help Spotify in targeted marketing campaigns, customer retention strategies, and optimizing service offerings.

Dataset Description

The dataset spotify_user_behavior_realistic_50000_rows.csv contains various attributes related to Spotify user behavior. Key features include:

user_id: Unique identifier for each user.
country: User's country of residence.
age: User's age.
signup_date: Date when the user signed up.
subscription_type: Type of Spotify subscription (e.g., Free, Premium Individual, Premium Duo, Premium Family).
subscription_status: Target variable indicating whether the user is 'Active' or 'Inactive'.
months_inactive: Number of months the user has been inactive.
inactive_3_months_flag: Binary flag indicating if the user has been inactive for 3 or more months.
ad_interaction: User's interaction with ads (Yes/No).
ad_conversion_to_subscription: Whether ad interaction led to subscription (Yes/No).
music_suggestion_rating_1_to_5: Rating of music suggestions (1-5).
avg_listening_hours_per_week: Average listening hours per week.
favorite_genre: User's favorite music genre.
most_liked_feature: User's most liked Spotify feature.
desired_future_feature: Feature desired by the user in the future.
primary_device: Primary device used for Spotify.
playlists_created: Number of playlists created by the user.
avg_skips_per_day: Average number of song skips per day.

The target variable is subscription_status.

Feature Engineering and Preprocessing

Feature Removal: The user_id column was dropped as it is a unique identifier and not relevant for prediction.
Date Feature Engineering: The signup_date column was transformed into numerical features:
- account_age_days: Number of days since signup until today.
- signup_month: Month of signup.
- signup_year: Year of signup. The original signup_date column was then dropped.
Data Type Categorization:
- skewed_col: ["music_suggestion_rating_1_to_5", "months_inactive", "inactive_3_months_flag", "playlists_created", "avg_skips_per_day"] - these were identified for power transformation.
- categorical_col: ["country", "subscription_type", 'ad_interaction', "primary_device", 'ad_conversion_to_subscription', "favorite_genre", "most_liked_feature", "desired_future_feature"] - these were identified for one-hot encoding.
- numerical_col: ["age", "avg_listening_hours_per_week", "account_age_days", "signup_month", "signup_year"] - these were identified for standard scaling.
Preprocessing Pipelines:
- skewed_pipe: Applies PowerTransformer (Yeo-Johnson method) to handle skewed numerical features.
- cat_pipe: Applies OneHotEncoder to convert categorical features into numerical format.
- num_pipe: Applies StandardScaler to standardize numerical features. These individual pipelines were combined into a ColumnTransformer (preprocessor).
Train-Test Split: The dataset was split into training and testing sets with a 80/20 ratio, using stratify=y to maintain the proportion of target classes.

Modeling Approaches Explored

The following classification models were explored for predicting subscription status:

Logistic Regression: A linear model for binary classification, often serving as a strong baseline.
- Without SMOTENC: Initial tuning performed using GridSearchCV on parameters like C, penalty, and solver.
- With SMOTENC: SMOTENC (Synthetic Minority Over-sampling Technique for Nominal and Continuous) was applied within a pipeline to handle potential class imbalance before applying Logistic Regression.
RandomForestClassifier: An ensemble learning method based on decision trees, known for its robustness and accuracy. Hyperparameter tuning was performed using RandomizedSearchCV.
GradientBoostingClassifier: Another powerful ensemble technique that builds trees sequentially, with each new tree correcting errors from previous ones. Hyperparameter tuning was performed using RandomizedSearchCV.
XGBoost (XGBClassifier): An optimized distributed gradient boosting library designed for speed and performance. Hyperparameter tuning was performed using RandomizedSearchCV.

Best Model: Logistic Regression (without SMOTENC, with hyperparameter tuning)

After evaluating all models, the Logistic Regression model, tuned with GridSearchCV and without the use of SMOTENC (due to similar performance across SMOTENC and non-SMOTENC approaches, favoring the simpler model), was selected as the best performing model. It achieved excellent metrics on the test set.

Hyperparameters:

The best hyperparameters for the Logistic Regression model were determined to be C=0.01, penalty='l2', and solver='lbfgs'.

Performance Metrics:

Accuracy: 0.9375
Precision (Active): 1.00
Recall (Active): 0.9258
F1 Score (Active): 0.9615

Classification Report:

              precision    recall  f1-score   support

      Active       1.00      0.93      0.96      8422
    Inactive       0.72      1.00      0.83      1578

    accuracy                           0.94     10000
   macro avg       0.86      0.96      0.90     10000
weighted avg       0.96      0.94      0.94     10000

The model shows perfect precision for the 'Active' class, meaning that when it predicts a user is active, it is always correct. The recall for the 'Inactive' class is 1.00, indicating that it correctly identifies all inactive users. However, the precision for 'Inactive' is 0.72, suggesting some active users are misclassified as inactive. Given the overall accuracy and F1-score for the 'Active' class, this model is highly effective.

Visualizations:

Confusion Matrix:

The confusion matrix shows that the model correctly predicted 7797 active users and 1578 inactive users. There were 625 active users incorrectly predicted as inactive, and 0 inactive users incorrectly predicted as active. This aligns with the high precision for 'Active' and high recall for 'Inactive'.

ROC Curve:

The Area Under the Receiver Operating Characteristic Curve (AUC-ROC) for the best model is approximately 0.962, indicating excellent discriminatory power between active and inactive users.

Precision-Recall Curve:

The Precision-Recall Curve illustrates the trade-off between precision and recall for different thresholds. The curve shows high precision even at relatively high recall values, especially for the positive class.

Model Saving and Loading

The best performing model (Logistic Regression with hyperparameter tuning, without SMOTENC) has been saved using joblib.

How to Save the Model:

import joblib
joblib.dump(grid, "model.pkl")

The grid object here represents the GridSearchCV fitted object, which internally stores the best estimator (the pipeline with the Logistic Regression model). The model was saved both locally and to Google Drive.

How to Load the Model:

To load the saved model and make predictions on new data:

import joblib

# Load the model from a local file
model = joblib.load("model.pkl")

# Or, if saved to Google Drive, load from there
# from google.colab import drive
# drive.mount('/content/drive')
# model = joblib.load("/content/drive/MyDrive/model.pkl")

# Assuming `new_data` is a pandas DataFrame with the same features as the training data
# predictions = model.predict(new_data)
# probabilities = model.predict_proba(new_data)

This README.md provides a comprehensive overview of the project, from data preparation to model selection and deployment, emphasizing the performance and utility of the chosen Logistic Regression model.

Summary:

Data Analysis Key Findings

Project Objective: The primary goal was to predict Spotify user subscription status (Active/Inactive) using a dataset of 50,000 user behavior records.
Feature Engineering Highlights:
- The user_id column was removed as it's not a predictive feature.
- The signup_date was transformed into account_age_days, signup_month, and signup_year to capture temporal patterns.
- Features were categorized and preprocessed using PowerTransformer for skewed numerical features, OneHotEncoder for categorical features, and StandardScaler for numerical features.
Modeling Approaches: Several classification models were explored, including Logistic Regression, RandomForestClassifier, GradientBoostingClassifier, and XGBoost, with hyperparameter tuning applied to each. SMOTENC was considered for class imbalance but ultimately not used in the best model.
Best Performing Model: A hyperparameter-tuned Logistic Regression model (without SMOTENC) was selected as the best performer, achieving strong results on the test set.
- The optimal hyperparameters were C=0.01, penalty='l2', and solver='lbfgs'.
Model Performance Metrics:
- Accuracy: 0.9375
- Active Class Metrics: Precision of 1.00, Recall of 0.9258, and F1 Score of 0.9615. This indicates that when the model predicts a user is 'Active', it is always correct.
- Inactive Class Metrics: Precision of 0.72, Recall of 1.00, and F1 Score of 0.83. This means the model successfully identified all 'Inactive' users, but some 'Active' users were misclassified as 'Inactive'.
- Overall Discriminatory Power: The Area Under the Receiver Operating Characteristic Curve (AUC-ROC) for the best model was approximately 0.962, indicating excellent ability to distinguish between classes.
Confusion Matrix Observations: The model correctly predicted 7797 'Active' users and 1578 'Inactive' users. Notably, 625 'Active' users were incorrectly predicted as 'Inactive', while zero 'Inactive' users were incorrectly predicted as 'Active'.
Model Persistence: The final best model has been saved as model.pkl using joblib for future deployment and inference.

Spotify User Behavior Prediction 🎧