Playground Series S5E8

Binary Classification with a Bank Dataset

Predicting whether bank clients will subscribe to term deposits based on demographic data, contact history, and economic indicators to optimize marketing campaigns.

Competition Rank

576 / 3,367

Percentile

Top 17%

Evaluation Metric

ROC-AUC

Participation

First 7 days

Problem Overview

This competition involved predicting customer responses to bank marketing campaigns - a classic imbalanced classification problem with significant business impact. The dataset included customer demographics, previous campaign interactions, economic context indicators, and contact details. Success in this task directly translates to more efficient marketing spend and better customer targeting.

Technical Approach

Advanced Feature Engineering: Created features capturing campaign recency, contact frequency patterns, and customer lifecycle indicators
Stratified K-Fold Validation: Implemented 10-fold stratified CV to handle class imbalance effectively
Hyperparameter Optimization: Used Optuna for systematic hyperparameter tuning across all models
Ensemble Architecture: Blended LightGBM, XGBoost, and CatBoost with optimized weights
Class Imbalance Handling: Applied scale_pos_weight and focal loss to address the minority class
Feature Selection: Removed multicollinear features and used SHAP for feature importance analysis

Key Insight

Despite only participating in the first 7 days of the competition, I achieved top 17% through efficient iteration and focused experimentation. The key was identifying that previous campaign outcomes were the strongest predictors, but combining them with economic indicators (employment variation rate, consumer confidence index) provided crucial context. This competition reinforced that domain understanding accelerates model development significantly.

Technology Stack

Python Pandas NumPy Scikit-learn LightGBM XGBoost CatBoost Optuna SHAP Imbalanced-learn

Lessons Learned

This was an incredibly fun competition that taught me the value of rapid prototyping and systematic experimentation. By structuring my workflow with version control, configuration management, and automated validation pipelines, I could iterate quickly even with limited time. The experience mirrors real-world data science where you often need to deliver results under time constraints.

From a DevOps perspective, I also set up a reproducible pipeline using Python scripts with configuration files (YAML), making it easy to track experiments and reproduce results - practices I now use daily in my role as a Data and DevOps Engineer.