🧠 Daily Study Log [2025-07-23]
Today’s focus was on idea generation, competition modeling, paper review and practical implementation, along with team discussion and feedback sharing.
A solid day of both solo and collaborative learning.

💡 Project Idea — Trust or Trick? Detecting Suspicious Movie Reviews through Behavioral Signals

Outlined a new NLP project designed to identify potentially deceptive or promotional movie reviews.
The model focuses on mismatches between star ratings and textual sentiment, along with reviewer metadata to detect anomalies.

Key elements:

Sentiment vs. star score inconsistency detection
Repetitive use of promotional keywords
Reviewer behavior analysis (e.g. single 5-star, no review history)
Clustering suspicious review bursts
Visualizations: heatmaps and review similarity graphs
Important considerations: privacy, fairness, false positives

📚 TOEIC Practice

Continued TOEIC study with a focus on consistency.
Practiced grammar (e.g. part 5), phrasal verbs, and reading sections.
Goal was to reinforce habits and maintain test readiness.

🏆 Competition — Electricity Usage Forecasting (DACON)

Tracked submission progress and improvements across multiple experiments:

No.	Description	Details	SMAPE
4	RandomSearchCV tuning (XGBoost)	Basic feature set, 3-fold CV	17.75976
5	Hold-out validation + Feature Engineering	Date-based split, temporal + building features + Voting ensemble	16.15700
6	Optuna tuning + Ensemble	Individually tuned XGB/LGBM/GBR with Optuna, then ensembled	11.33852

Notes:

RandomSearchCV improved local performance but showed overfitting signs
Submission 5 introduced hold-out strategy and meaningful temporal features
Submission 6 used per-model Optuna tuning → major improvement on public board

Next plans:

Try StackingRegressor instead of voting
Train models per building type (e.g. hospitals vs. offices)
Introduce lag and rolling features
Consider time-based weighting for recent samples

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

Completed the full reading of the paper and prepared for implementation.
Key takeaways:

Semi-supervised 3D pose estimation achieves ~47mm MPJPE on Human3.6M
Genre classification improved most when combining 3D pose and inferred motion
Future direction: motion synthesis, unsupervised pretraining

🧪 Practice — 국민체조 Pose Estimation

Tested 2D skeleton extraction from a YouTube video of 국민체조 (National Calisthenics).
Extracted frames and processed them using pose estimation code.
Uploaded the notebook: 2D_Pose_Feature_Builder.ipynb
This forms the foundation for further experiments in skeleton similarity and feedback systems.

🧠 Paper Study Group — Feedback & Discussion

Shared the 국민체조 code implementation with the group; feedback was positive
Discussed practical applications of the code in motion feedback
Compared different papers on pose-based evaluation and matching metrics
Explored potential for extending to healthcare and user-facing feedback systems

✅ TL;DR

📍 Idea: Designed “Trust or Trick?” suspicious review detection project
📍 TOEIC: Practiced grammar, reading, and reviewed idioms
📍 Competition: Logged 3 key submissions; Optuna + ensemble was the strongest
📍 Paper: Fully read HDVR paper and prepared to implement
📍 Practice: Ran 국민체조 2D skeleton estimation
📍 Study Group: Shared code and discussed paper-based extensions