🛍️ E-Commerce Customer Retention Classification

🔗 Objective

This project focuses on building a supervised machine learning classification model to predict whether a customer will be retained (subscribe) or churn (not subscribe) based on their demographic, transactional, and behavioral features. The insights derived from this model aim to help e-commerce businesses tailor loyalty programs and marketing strategies to improve retention and revenue.

---

Dataset Overview

The dataset includes features categorized as:

• Demographics: Age, Gender, Location
• Purchase Behavior: Item Purchased, Category, Purchase Amount (USD), Previous Purchases
• Customer Engagement: Review Rating, Subscription Status, Discount Applied, Promo Code Used
• Shipping & Payment: Shipping Type, Payment Method

---

📊 Exploratory Data Analysis (EDA)

• Checked for duplicates, missing values, and outliers
• Explored numerical statistics using .describe()
• Used pairplots and correlation analysis (Pearson and Spearman) to check linear and monotonic relationships

Key Findings:

• No strong linear or monotonic relationship between numerical features
• Subscription status is imbalanced: 73% No, 27% Yes
• Categorical analysis showed:
  • Promo code usage and frequent purchases correlate with higher subscriptions
  • Discounts and free shipping are not strong incentives for subscription
  • Clothing and accessories had higher non-subscription rates

---

Feature Engineering & Preprocessing

• Categorical features: One-hot encoded
• Ordinal features: Ordinal encoded with custom order (e.g., Size, Season, Frequency)
• Binary features: Ordinal encoded
• Numerical features: Standard scaled
• Built a ColumnTransformer pipeline for clean preprocessing

---

Models Trained

Six classification models were evaluated using 5-fold cross-validation:

• Dummy Classifier (baseline)
• Logistic Regression ✅ Best Model
• SVM (RBF Kernel)
• Random Forest
• KNN
• Decision Tree

Evaluation Metrics:

Model	Accuracy	F1 Score	Recall	Precision	AUC
Logistic Regression	83.43%	76.52%	100%	61.97%	0.91
SVM (RBF)	83.40%	76.48%	100%	61.92%	-
Random Forest	82.08%	70.92%	81.00%	63.10%	-
Decision Tree	80.32%	62.65%	61.29%	64.22%	-
KNN	79.74%	62.23%	61.88%	62.63%	-
Dummy Classifier	61.15%	29.89%	30.64%	29.23%	-

---

Handling Class Imbalance & Evaluation Strategies

Several resampling and adjustment techniques were applied to improve model performance on the imbalanced dataset (73% non-subscribers vs. 27% subscribers):

Evaluation of Techniques

• Original Model:

  • High accuracy and recall
  • Lower precision → Model predicts many false positives
  • AUC shows good class separation

• Balanced Resampling:

  • Perfect recall (100% of subscribers identified)
  • Slightly lower precision
  • Strong AUC and high F1 score → balanced and effective

• Class Weight Adjustment:

  • Poor recall and precision
  • Very low F1 score
  • AUC significantly lower than other methods → not suitable

• SMOTE (Synthetic Minority Over-sampling Technique):

  • Perfect recall
  • High F1 score
  • AUC improves → model better distinguishes classes

• Undersampling:

  • Perfect recall
  • High F1 score
  • Strong AUC, similar to SMOTE

Conclusion:

• SMOTE, undersampling, and the balanced strategy are the most effective techniques.
• These methods achieved perfect recall and high F1 scores.
• Class weight adjustment underperformed and is not recommended for this dataset.

---

ROC & AUC

• AUC = 0.91 for Logistic Regression
• Model performs significantly better than random guessing (baseline)

Confusion Matrix

Visualized to better understand the prediction errors.

---

Conclusion

• Logistic Regression is the best performing model in terms of accuracy, recall, and AUC.
• Recall = 1.00, which means the model catches all retained (subscribed) customers.
• Model is ideal for business deployment where customer retention is key.

---

Tools & Libraries

• Python (Pandas, NumPy, Seaborn, Matplotlib, Plotly)
• Scikit-learn
• imbalanced-learn (SMOTE, RandomUnderSampler)

---

Future Improvements

• Apply advanced models (XGBoost, Gradient Boosting)
• Analyze feature interactions

---

🤝 About the Author

Shiva Dorri
Data Scientist | Background in Accounting & Finance | Passionate about solving business problems with machine learning and data analysis.

"I believe that real impact comes from turning data into decisions."

---

Techniques used: Logistic Regression SMOTE UnderSampling Classification Customer Retention E-commerce Supervised Learning Feature Engineering Imbalanced Data Scikit-learn