Customer Churn Prediction

Project Overview

This project focuses on predicting customer churn using machine learning techniques. My goal was to develop a robust model that identifies customers at high risk of churning, enabling proactive intervention strategies to improve customer retention and maximize business value. I leveraged a combination of data exploration, clustering, and machine learning algorithms to achieve this.

1. Data Exploration & Preprocessing

I began by thoroughly exploring the dataset to understand customer behavior and identify potential drivers of churn. Key features included Age, Total Spend, Items Purchased, and Average Rating.

Distribution of customer age before outlier handling.

Distribution of customer total spend before outlier handling.

Distribution of items purchased before outlier handling.

Distribution of customer average rating before outlier handling.

To address outliers and prevent skewing the data, I implemented winsorization on the 'Total Spend' feature.

Distribution of Total Spend after winsorization.

2. Clustering

Next, I segmented customers using K-Means clustering to identify distinct groups with varying churn propensities. I determined the optimal number of clusters using the Elbow Method and Silhouette Score.

Elbow method for determining optimal number of clusters.

Silhouette score for evaluating cluster quality.

Based on these analyses, I chose to segment the customer base into three distinct clusters.

# Code Snippet (Example - Replace with your actual code)
from sklearn.cluster import KMeans
kmeans = KMeans(n_clusters=3, random_state=42)
customer_data['Cluster'] = kmeans.fit_predict(customer_data[features])

The characteristics of each cluster are summarized below:

Cluster	Avg. Age	Avg. Total Spend	Avg. Items Purchased	Avg. Average Rating
0	39.36	473.48	8.49	3.33
1	29.92	1310.97	17.62	4.68
2	31.56	748.43	11.66	4.05

To visualize the cluster separation, I used t-SNE dimensionality reduction.

t-SNE visualization of customer clusters t-SNE visualization of customer clusters.

The cluster sizes were as follows:

Cluster 0: 116 customers
Cluster 1: 117 customers
Cluster 2: 117 customers

3. Churn Prediction

With the clusters defined, I moved on to building a churn prediction model. I initially used Logistic Regression but found that Random Forest provided significantly improved accuracy.

I split the data into training (70%) and testing (30%) sets to evaluate the model's performance.

Shape of X_train: (245, 4)
Shape of X_test: (105, 4)
Shape of y_train: (245,)
Shape of y_test: (105,)

Logistic Regression Results:

Accuracy: 0.9143
AUC: 0.9907
Classification Report:
              precision    recall  f1-score   support

           0       0.94      0.95      0.94        77
           1       0.85      0.82      0.84        28

    accuracy                           0.91       105
   macro avg       0.89      0.88      0.89       105
weighted avg       0.91      0.91      0.91       105

Random Forest Results:

Random Forest Accuracy: 1.0
Random Forest AUC: 1.0
Classification Report:
              precision    recall  f1-score   support

           0       1.00      1.00      1.00        77
           1       1.00      1.00      1.00        28

    accuracy                           1.00       105
   macro avg       1.00      1.00      1.00       105
weighted avg       1.00      1.00      1.00       105

The Random Forest model achieved perfect accuracy and an AUC of 1.0, demonstrating its strong predictive power.

Feature Importance:

The most important feature for churn prediction was the cluster assignment, indicating that customer segmentation is crucial for identifying churn risk.

Feature Importance:
Cluster: 0.420
Items Purchased: 0.288
Age: 0.104
Total Spend: 0.035

4. Intervention Strategies & Cost-Benefit Analysis

Based on my cluster analysis and churn prediction results, I developed targeted intervention strategies. I focused on Cluster 0, which exhibited a high churn rate of 71.6%.

Intervention: Personalized Communication (Cluster 0)

• Assumptions:

  • Cost per personalized email: $0.10
  • Number of customers in Cluster 0: 58
  • Current churn rate in Cluster 0: 71.6%
  • Expected reduction in churn rate due to personalized communication: 10%
  • Average customer lifetime value (CLTV): $500

• Cost Calculation:

  • Total cost of sending personalized emails to Cluster 0: 58 customers * $0.10/email = $5.80

• Benefit Calculation:

  • Number of customers retained due to the intervention: 58 customers * 10% churn reduction = 5.8 customers
  • Revenue saved by retaining these customers: 5.8 customers * $500/customer = $2900

• Cost-Benefit Analysis:

  • Total Cost: $5.80
  • Total Benefit: $2900
  • Net Benefit: $2894.20
  • Return on Investment (ROI): 49834%

• Interpretation:

  • The cost-benefit analysis indicates that personalized communication for Cluster 0 is a highly effective and profitable churn reduction strategy.

Conclusion

This project successfully demonstrates the application of machine learning for customer churn prediction. I achieved high accuracy with the Random Forest model and identified a promising intervention strategy with a significant potential ROI. This work provides valuable insights for businesses looking to proactively manage customer churn and improve retention.

Future Work

• Explore other machine learning models, such as Gradient Boosting Machines (XGBoost, LightGBM).
• Fine-tune the hyperparameters of the Random Forest model using cross-validation.
• Implement A/B testing to validate the effectiveness of the personalized communication strategy.
• Develop a real-time churn prediction system.

Dependencies

• Python
• pandas
• scikit-learn
• matplotlib
• seaborn