LG Aimers 5th: Product Defect Detection with Stacking and SMOTE-ENN

Summary

• Original article title, translated: "LG Aimers 5th: Product Defect Detection with Stacking and SMOTE-ENN." Only the filename's unsupported colon was replaced. Published 2024-10-25 · slug lg-aimers-5th-comprehensive-retro · category "Projects, hackathons, and competitions / ML and data competitions" · source article recorded in Korean and English.
• A retrospective on the fifth LG Aimers "Product Defect Detection" competition. The central challenges were severe class imbalance, with defective products below 1% of the total, and evaluation using F1-Score.
• I used SMOTE-ENN hybrid sampling, reporting an eight-percentage-point precision improvement over SMOTE alone; stacking with LGBM, XGBoost, and CatBoost base models and a logistic-regression meta-model; and stratified five-fold out-of-fold training.
• The result was in the top 5.5% among 1,400 teams, narrowly missing the final round.
• The retrospective identified three shortcomings: insufficient feature engineering, weak experiment management that led to repeated experiments, and poor time allocation. The plan for the next competition was to reserve separate periods for EDA and feature creation.

Mentioned

• LG Aimers 5th cohort: the competition and training program discussed in the article.
• LG AI Research: the organizer of LG Aimers.

Connections

• Writing · Credentials
• Atomic notes: Without experiment management, the same experiments get repeated; schedule EDA and feature creation separately · Severe imbalance can benefit from combined sampling; SMOTE-ENN improved precision by eight percentage points over SMOTE alone
• Concepts: CatBoost · F1-Score · LightGBM · SMOTE-ENN · Stacking ensembles · XGBoost · i18n and multilingual support · Cross-validation · Running a technical blog · Logistic regression · Ensembles · Anomaly detection · Class imbalance · Feature engineering · Hyperparameter tuning