🧠 Daily Study Log [2025-08-05]
Took it a bit slower today due to low energy, but still made progress on idea generation, SQL fundamentals, competition experiments, and continued the causal inference paper review.

💡 Project Idea — AI Fatigue Journal

Proposed a self-insight tool that estimates mental fatigue levels based on free-form text like journals or notes.
Goes beyond simple writing — it analyzes patterns to detect signs of fatigue and visualizes trends over time.

Concept:

• Input: daily journals, memos, or messages
• Analysis: fatigue-related keywords, emotional tone, repetitive expressions
• Output: fatigue score, trend visualization, language habit feedback

🔗 View full idea

🧪 Competition — Electricity Usage Forecasting (DACON)

Continued with enhancements to the segmented VotingRegressor architecture:

25. Outlier Correction using IQR

• Applied IQR filtering on VotingRegressor predictions to reduce extreme noise
  📊
• Local SMAPE: 11.6907
• Public LB: 15.1169

26. Added Nonlinear Interaction Features

• Kept the Time × Building Type segmented model
• Added new features like Season × BuildingType, Temperature × Area, Humidity × CoolingArea
  📊
• Local SMAPE: (Pending)
• Public LB: 15.0829

27. Feature Selection with Top 20 Important Features

• Selected top 20 features based on importance
• Aimed to reduce model complexity and improve generalization
  📊
• Local SMAPE: (Pending)
• Public LB: (Pending)

🧮 SQL Study — DDL (Data Definition Language)

Studied the DDL part of SQL today.
Key focus:

• CREATE, ALTER, DROP, TRUNCATE
• DDL statements define and modify schema
• Usually auto-committed → irreversible if done wrong

📖 Paper Review — DoWhy: A Library for Causal Inference

Continued reviewing the DoWhy framework.
Today’s focus was on the importance of explicitly declaring assumptions and formalizing causal beliefs through DAGs.

Key takeaways:

• “Assumptions are the model.”
• Causal inference is about interventions and explanation, not just prediction
• The 4-step approach (Model → Identify → Estimate → Refute) provides a rigorous workflow

📂 View notes and code

✅ TL;DR

📍 Idea: AI Fatigue Journal — text-based mental fatigue estimation and visualization
📍 Competition:

• 25: Outlier smoothing using IQR (SMAPE: 15.1169)
• 26: Added nonlinear interaction features (SMAPE: 15.0829)
• 27: Feature selection with top 20 variables (TBD)
  📍 SQL: Studied DDL (CREATE, ALTER, DROP, etc.)
  📍 Review: DoWhy — structured causal reasoning through assumptions and DAGs