🧠 Daily Study Log [2025-06-30]
Today’s work centered on two fronts: pushing SCU_Competition to new heights through classic feature optimization and restructuring my GitHub Pages to enhance portfolio visibility and navigation.

📊 SCU_Competition — Submission 31–33

Focus: Maximizing AUC with simpler, classic structures + filtered feature interactions
Model: Optuna-tuned LGBMClassifier + StratifiedKFold CV + EarlyStopping

Strategies Applied:

• Submission 31: Strategic feature selection + Optuna + interaction terms → Kaggle AUC 0.8935
• Submission 32: More refined Optuna tuning with extended trials → Kaggle AUC 0.8861
• Submission 33: CV-based tuning + conservative EarlyStopping → Kaggle AUC 0.8808

Key Observations:

• Submission 31 proved that classic + clean interaction features are still top-tier
• Submission 32 reached the highest local AUC (0.9057) but saw slight generalization drop
• Submission 33 provided a safer tuning route, avoiding overfitting via tighter validation

✅ Takeaways

• Conservative modeling remains strong when paired with selective feature engineering
• EarlyStopping helps regularize even aggressive Optuna-based trials
• Interaction features must be filtered for actual impact, not just volume

🛠 GitHub Pages Restructuring

Goal: Improve readability and navigation of my portfolio blog
Actions Taken:

• Refined category and folder structure (/paper_review, /projects, /study_log)
• Added visual spacing, layout clarity, and more intuitive headings
• Fixed post redundancy and performed manual syncing with local git repo

Outcome:

• Blog now serves as a cleaner, more organized public-facing tech portfolio
• Easier navigation for professors, recruiters, or collaborators

🎯 Next Steps

• Submission 34: Combine clean modeling with SHAP-based filtering
• Integrate auto feature selectors (e.g., Boruta, RFECV) into baseline
• Add permanent sidebar or tag system in blog for quicker topic access

✅ TL;DR

📍 SCU: Classic + smart feature combos still outperform complex stacks
📍 Optuna: More trials ≠ better Kaggle results — balance needed
📍 GitHub: Clean portfolio blog is now easier to browse and share