B2B marketing-qualified lead (MQL) prediction

📋 Project Overview

This project aims to develop machine learning models to predict sales success. Through various experiments with different models and optimization techniques, we achieved significant improvements in prediction performance.

🔬 Experiments & Results

Experiments	Code	Score
Decision Tree	Decision_Tree.ipynb	0.5751
Ensemble	Ensemble_0.5751.ipynb	0.5751
XGBClassifier & Optuna	XGBClassifier&Optuna_0.597.ipynb	0.5922
Optuna2	Optuna2_0.597.ipynb	0.597
XGBClassifier	XGBClassifier_0.601.ipynb	0.601
Ensemble & Optuna	Ensemble&Optuna_0.5922.ipynb	0.5922
Final Model	Final_0.62.ipynb	0.62

🛠 Data Processing & Feature Engineering

Data Preprocessing Steps

1. Label Encoding for categorical variables:

• customer_country
• business_area
• business_unit
• customer_type
• enterprise
• customer_job
• inquiry_type
• product_category
• and more...

2. Missing Value Treatment

• Numerical columns: Filled with median values
• Categorical columns: Filled with 'missing'
• Strategic variables (id, it, idit): Filled with 0
• com_reg_ver_win_rate: Median value
• ver_win_rate_x: Median value
• ver_win_ratio_per_bu: Median value

3. Feature Selection

• Removed columns: id_strategic_ver, it_strategic_ver, idit_strategic_ver, lead_desc_length
• Dropped: customer_type, historical_existing_cnt, product_subcategory, product_modelname
• Removed: expected_timeline, business_subarea

4. Imbalanced Data Handling

• Retained all positive samples (is_converted = 1)
• Undersampled negative cases to 25,000 samples

🤖 Model Development

1. Base Models

• Decision Tree Classifier (Score: 0.5751)
• XGBClassifier (Score: 0.601)
• Ensemble Models (Score: 0.5751 ~ 0.5922)

2. Advanced Approaches

• Ensemble Strategy

• Combined multiple models with weighted voting

• Weights distribution: [0.1, 0.1, 0.1, 1.2]

• Models included:

  • LogisticRegression
  • RandomForestClassifier
  • GradientBoostingClassifier
  • XGBClassifier

• Hyperparameter Optimization

• Utilized Optuna for automated parameter tuning

• Optimized parameters for each base model

• Enhanced ensemble performance through optimal weights

3. Final Model (Score: 0.62)

• Best performing model combining all optimizations
• Achieved through careful feature selection and model tuning