Customer Churn Prediction

📌 Project Workflow

1. Data Cleaning & Preprocessing

• Removed irrelevant columns and handled missing values.
• Converted Yes/No categorical columns into bool type.
• Reduced memory usage by approximately 80%, making the dataset more efficient for modeling.

2. Feature & Target Definition

• X (features): all columns except the target.
• y (target): the Churn column.

3. Train/Test Split

• Used train_test_split to split the dataset into training and testing sets.
• Applied stratify=y to maintain the same distribution of churn labels in both sets.

4. Pipeline Creation

• Built a pipeline with a ColumnTransformer and a classifier.
• Preprocessing:
  • OneHotEncoder for categorical features.
  • StandardScaler for numeric features.

5. Initial Model Benchmarking (GridSearchCV)

• Evaluated 4 classifiers with 5-fold cross-validation:
  • RandomForestClassifier
  • LogisticRegression (liblinear)
  • GradientBoostingClassifier
  • Support Vector Machine (SVM)
• Applied class_weight="balanced" (except GradientBoosting) due to imbalanced churn classes.
• Optimized for F1 Score since accuracy alone would be misleading.
• Result: Logistic Regression achieved the best performance in the initial stage.

6. Model Evaluation

• Reported metrics: Accuracy, Precision, Recall, and F1 Score.
• Visualized results for better interpretability.

7. Feature Selection (Feature Importance)

• Extracted the top 8 most influential features.
• Removed less relevant features to improve efficiency and reduce noise.

8. Second Round of Model Tuning

• Re-ran preprocessing and GridSearchCV on the reduced feature set.
• Result: RandomForestClassifier (200 estimators, max depth = 8) delivered the best performance.

9. Final Comparison & Insights

• GradientBoostingClassifier had the longest average training time.
• Between Logistic Regression and Random Forest:
  • Random Forest required more training time.
  • But achieved ~5% higher F1 Score.
• Final Takeaway: RandomForestClassifier provided the best trade-off between accuracy, robustness, and performance.

---

📊 Results

• Baseline Best Model: Logistic Regression (liblinear)
• Final Best Model after Feature Selection: RandomForestClassifier (200 estimators, max_depth=8)