🧠 Daily Study Log [2025-08-02]
From SQL theory to public leaderboard tuning — today was about deepening fundamentals and refining experimental pipelines.

🧪 Competition — Electricity Usage Forecasting (DACON)

Three new experiments focused on segmentation and weighted ensemble tuning.

19. Model Per Building Type

Trained separate models for each building type to better reflect distinct patterns.

→ Public LB: 76.7876 (aggregated submission)

20. Weighted Voting Regressor (Optuna-Tuned)

Applied manual weights (XGB: 0.5, LGBM: 0.3, GBR: 0.2) after tuning each model individually with Optuna.

• Local: 15.5039
• Public LB: 12.4674

21. Time-Split Voting Models

Segmented data by time of day:

• Night: 0–8h
• Day: 9–18h
• Evening: 19–23h

Trained separate VotingRegressor for each to reflect different load patterns.

• Local: 13.4856
• Public LB: 11.2358 ✅ Best so far

📝 SQL Study

Focused on TCL (Transaction Control Language) and DCL (Data Control Language):

• COMMIT / ROLLBACK / SAVEPOINT
• GRANT / REVOKE

💡 Project Idea — AI Pottery Assistant

Proposed a creative concept: an intelligent assistant that analyzes pottery shapes, designs, and user prompts to assist in creative ceramic design.

🔗 View full idea

📄 Paper Review — Explainable Skeletal Assessment for Martial Arts

✅ Finished Day 4: Experiments, Results, and SHAP-based interpretation.

Key Takeaways:

• Strong performance on XSQ and PBB datasets (even outperforming experts in MAE)
• SHAP visualizations provide local/global feedback for motion quality
• Ablation shows alignment improves performance across models
• Final reflection: effective, interpretable motion scoring pipeline

🔗 View summary

🧪 Personal Project — bass_seeker

Moved away from Spotify API and focused on audio-feature-based recommendation using .wav data.
Tested extraction pipeline using librosa and spectral centroid/bandwidth.
📍 Project: bass_seeker

🤝 Research Club — Taekwondo Poomsae Evaluation

Set new direction: RLHF-based quantitative evaluation of Taekwondo forms.

Highlights:

• Pose-based + Similarity + Human Feedback (RLHF)
• Goal: Not just what the score is — but why it was given
• Core tech: DTW alignment, SHAP feedback, reinforcement learning

📌 Read the full summary (in local repo)

✅ TL;DR

📍 Idea: AI Pottery Assistant — creativity aid for ceramic design
📍 SQL: TCL & DCL studied
📍 Competition: 3 advanced ensemble experiments submitted (Best SMAPE: 11.2358)
📍 Paper: Martial Arts scoring pipeline review completed
📍 Projects: bass_seeker progressed — audio-based pipeline working
📍 Study Group: RLHF-based Taekwondo scoring system planning initiated