Machine Learning Binary Classification Project

This project solves a binary classification problem using a dataset of 9,000 rows and 11 features (numerical and categorical). The goal is to predict a binary target variable (0 or 1) through preprocessing, feature engineering, model building, and explainability techniques.

Data Preprocessing

• Removed outliers in one step using a custom trimming function across all numeric columns.
• Normalized numeric features with MinMaxScaler to scale values between 0 and 1.
• Explored feature-target relationships with:
  • Matrix boxplots and scatter plots (catplot)
  • Correlation matrix
  • T-tests and Chi-square tests for feature relevance

Feature Engineering

• Retained key features: feature_1, feature_2, feature_3, feature_4
• Removed weak features: feature_5 to feature_8
• Converted category_1 and category_2 into numerical values using Label Encoding
• Created feature pairs (mean, sum, square, product), and selected the most correlated with the target

Machine Learning Models

Implemented and evaluated the following models:

• Decision Tree
• Random Forest
• Gradient Boosting

Each model was assessed for:

• Accuracy
• Precision / Recall
• Visual performance evaluation

Model Explainability

Used SHAP and LIME for model interpretability:

• Identified features with the most influence on predictions
• Justified decisions made by complex models
• Reduced model complexity by excluding less impactful features

Business Use Case: Marketing Optimization

Using SHAP and LIME for business alignment (e.g., marketing agency):

• Found top conversion factors (e.g., page views, discounts)
• Identified the most effective channels (Instagram > Google)
• Understood why certain users don’t convert (e.g., short session, no discount)

Applications:

• Optimize marketing budget by removing ineffective channels
• Personalize ads based on feature influence
• Improve UX by targeting pain points

Technologies Used

• Python
• Pandas, NumPy
• Seaborn, Matplotlib
• Scikit-learn
• SHAP, LIME
• Jupyter Notebook

Conclusion

This project demonstrates the complete machine learning pipeline—from data cleaning to feature engineering, model evaluation, and explainability—with direct applications to business decision-making.