🧠 Daily Study Log [2025-07-04]
Today’s work included structured experimentation in the SCU_Competition, SQL fundamentals study, and implementation + review of MobileNetV2 architecture.

📊 SCU_Competition — Submission 51–55

Focus: Feature pruning, voting weight tuning, correlation filtering, and stacking re-evaluation
Goal: Further refine ensemble efficiency and eliminate redundant signals

Highlights:

• Submission 51: VotingClassifier (7:2:1) – removed 3 low-impact features from 50
  → CV AUC: 0.8833 / Kaggle AUC: 0.8950
• Submission 52: LGBMClassifier only – no tuning, just raw fit on current features
  → CV AUC: 0.8798 / Kaggle AUC: 0.8879
• Submission 53: Removed bottom 10% of features by importance
  → CV AUC: 0.8828 / Kaggle AUC: 0.8951
• Submission 54: Removed correlated features (threshold=0.95)
  → CV AUC: 0.8835 / Kaggle AUC: 0.8950
• Submission 55: Switched to StackingClassifier (CV AUC fell slightly)
  → CV AUC: 0.8780 / Kaggle AUC: 0.8841

Next Ideas:

• Try LGBM as meta-model in stacking
• Simplify cluster usage and feature combinations

💾 SQL Practice — Basic Data Modeling

What I Learned:

• Key concepts: Entity, Attribute, Domain, Relationship
• Differentiated between strong vs. weak entities
• Covered examples like Student–Course (M:N), Employee–Department (1:N)
• Understood ERD design, PK/FK logic, and how entities convert to tables

Reflection:

• This foundation is crucial before jumping into JOINs or queries
• Visualizing relationships helps in schema planning for real-world datasets

📄 Paper Review — MobileNetV2 (Day 4)

Focus: Final architecture summary + writing complete implementation in PyTorch

Today’s Progress:

• ✅ Implemented InvertedResidual block (expand → depthwise → linear projection)
• ✅ Integrated optional residual connections
• ✅ Validated tensor flow via forward pass
• 📁 Notebook: MobileNetV2_Architecture_Implementation.ipynb

Insight:

• ReLU6 is carefully placed only before projection
• Linear bottleneck structure ensures compressed, lossless feature flow
• Residual connections are applied only when stride=1 and channel dims match

✅ TL;DR

📍 SCU: Voting weight tuning and correlation-based feature filtering both helped improve model clarity
📍 SQL: Learned the essence of data modeling — entities, attributes, relationships, and schema design
📍 CV: Successfully implemented MobileNetV2 core block with custom PyTorch code and tested it
📍 Next: Try Neural Style Transfer + continue SQL (JOINs) + MobileNetV2 vs ShuffleNetV2 paper comparison