🧠 Daily Study Log [2025-08-01]
A well-balanced day of idea generation, competition modeling, project transition, and academic reading.

💡 Project Idea — Sunburn Simulator

Proposed a playful yet science-backed idea to estimate sunburn risk based on skin tone, UV index, and exposure time.

🔗 View full idea

Key Points:

• Input: skin type, UV index, time outdoors
• Output: estimated time to sunburn
• Practical use in skincare or outdoor planning
• Potential for app with UV-based alert system

📝 TOEIC Practice

Studied:

• Listening: Part 3 dialogue comprehension
• Reading: Passage interpretation and vocabulary review

Keeping momentum in test prep with consistent focus.

🏆 Competition — Electricity Usage Forecasting (DACON)

Ran several submissions using different configurations of VotingRegressor combining XGBoost and LightGBM.

Recent Experiments:

1. Basic Voting ensemble with XGBoost + LightGBM (equal weights)
     → Local: 15.7148 | Public LB: 14.6649768819

2. Manually tuned weights (0.6 for XGBoost, 0.4 for LightGBM)
     → Local: 15.6376 | Public LB: 14.6108526927

3. Auto-tuned weights using Optuna for the Voting ensemble
     → Local: 15.5960 | Public LB: 14.5501608484
      (best xgb_weight ≈ 0.71)

Highlights:

• Extensive feature engineering: datetime, weather, and building-related variables
• Applied LabelEncoder + OneHotEncoder to categorical building types
• Voting ensemble trained and validated using hold-out from August
• Tracking SMAPE on validation and public leaderboard

Next Steps:

• Segment models by building type
• Experiment with lag-based temporal features
• Try feature selection or add a third model (e.g., GBR) with weighted voting

📄 Paper Review — Explainable Skeletal Assessment for Martial Arts

✅ Finished Day 3 — Continued analysis of motion scoring and SHAP interpretation
🔗 View summary

Key Takeaways:

• End-to-end: alignment → scoring → explainability
• DTW + skeleton normalization central to scoring
• SHAP used for joint-wise contribution interpretation
• Planning to implement a 2D-based scoring prototype next

🔁 Project Transition

✅ PoseSyncEvaluator

• Fully implemented and launched locally with Streamlit
• All frame data handled and documented in GitHub

🚀 bass_seeker

• New project initialized for sound/genre analysis
  🔗 Repo Link

🤝 Research Club — New Topic: Taekwondo Poomsae Evaluation

Discussing a new theme for the research group — pose-based evaluation of Taekwondo forms.

Related Papers:

• Vision-Based Recognition of Taekwondo Poomsae Using Deep Learning, IEEE Access, 2020
• Automatic Skill Evaluation in Martial Arts using 3D Pose Estimation, CVPRW 2021
• Pose-based Action Recognition for Taekwondo Movements, Sensors, 2022
• Deep Learning-Based Action Quality Assessment in Taekwondo, Applied Sciences, 2023

Focus: skeleton-based form alignment, scoring, and explainability.

✅ TL;DR

📍 Idea: Sunburn Simulator — UV + skin type → burn-time estimation
📍 TOEIC: Listening Part 3 + Reading drills
📍 Competition: VotingRegressor running (XGBoost + SVR)
📍 Paper: SHAP + DTW scoring method reviewed
📍 Projects: PoseSyncEvaluator done → bass_seeker started
📍 Study Group: Brainstorming Taekwondo-based evaluation research