🧠 Daily Study Log [2025-07-22]
A productive day of multi-threaded progress: competition tuning, paper writing, new ideas, SQL theory, and project housekeeping.
Some directions are becoming clearer — especially around personalization and explainable modeling.

📄 Paper Review — Unsupervised 3D Pose Estimation for Hierarchical Dance Video Recognition

Wrapped up Day 4 of the paper review.
Focused on the experiments, performance results, and conclusion/future work.

Key takeaways:

• Semi-supervised 3D pose estimation achieves near SOTA accuracy (MPJPE ~47mm on Human3.6M)
• Genre classification improves most when combining 3D pose + inferred movements (F1 = 0.86)
• Future direction includes fully unsupervised training and motion-based dance synthesis

This paper continues to stand out for its modularity and practical evaluation pipeline.

💡 Project Idea — Travel With Whom: Context-Aware Review Mining

Explored a new NLP project idea centered around classifying reviews based on travel companion context.
This could enrich travel recommendation systems by understanding whether a place is good for families, couples, friends, or solo travelers.

Key components:

• BERT-based context detection (e.g., “traveled with my kids”)
• Movement of sentiment by group type
• Visualization layer showing group-specific ratings
• Optional recommender layer based on travel companion profile

Promising direction for both academic application and product design.

📊 Competition Progress — Electricity Usage Forecasting (DACON)

Made significant advances on model tuning and ensemble strategies:

Strategies applied:

1. RandomSearchCV tuning
   • Models: XGBoost, LGBM, GBR
   • Combined via soft voting
   • Local SMAPE: 8.88 / Public SMAPE: 17.75
2. Feature engineering + ensemble with hold-out validation
   • Added building features and time-based variables
   • Hold-out split based on latest timestamps
   • Local SMAPE: 17.16 / Public SMAPE: 16.15
3. Optuna-based tuning + final ensemble (in progress)
   • Tuning XGBoost, LightGBM, GBR separately
   • Using best params for final soft-voting
   • Hold-out split and building features applied

🧠 SQL Practice — Subqueries

Reviewed SQL subquery patterns:

• Nested SELECTs inside WHERE/IN clauses
• Correlated subqueries for row-wise conditions
• Filtering based on aggregated comparisons

Trying to solidify SQL foundations before I dive deeper into window functions and analytics.

📚 Paper Group — Direction Setting & Exploration

As part of leading our research reading group, I:

• Refined the theme: Pose-based motion analysis for dance/action/health
• Revisited the SEVENTEEN paper on matching rate of synchronized dance (2D + 3D evaluation)
• Skimmed a few more papers related to pose similarity and matching metrics

Planning to map these to future projects involving performance feedback and skeleton tracking.

🗂 GitHub Cleanup

Wrapped up the day with some repository organization:

• Updated READMEs
• Added clearer folder structures
• Cleaned up old notebooks
• Logged progress from DACON + pose-based work

Trying to maintain a clean portfolio as projects evolve.

✅ TL;DR

📍 Paper: Finished experiment/conclusion section of HDVR pose paper
📍 Idea: Designed “Travel With Whom” context-aware review system
📍 Competition: Tuned and ensembled multiple models (RandomSearch, Optuna, Hold-out)
📍 SQL: Focused study on subqueries
📍 Study Group: Set direction & reviewed SEVENTEEN pose paper
📍 GitHub: Final cleanup and consolidation