🧠 Daily Study Log [2025-08-18]
Worked on a new idea project, practiced SQL exam problems, advanced competition experiments with blending stability, reviewed experiments & results in the few-shot adaptation paper, continued development of the Bass Seeker project, and structured the novelty pipeline for the arm wave quality assessment paper.

💡 Idea Generation — Fit 2 Give

Proposed a concept combining fitness tracking and charitable giving, where calories burned or workout milestones trigger donations to social causes.
The idea links personal health motivation with social contribution.

🔗 View idea

🏆 Competition — Electricity Forecasting

Refined building-type–specific blending strategies to mitigate overfitting:

No.	Description	Local SMAPE	Public LB SMAPE
39	Addressed overfitting: tuned XGB/LGBM via RandomizedSearchCV for each building type, applied 7-Fold OOF, computed optimal weights per fold and averaged them, then retrained with 5 different seeds for seed averaging. (best)	4.065438	7.0021252141
40	Integrated flexible folds, negative prediction clipping, standardized LGBM params, adaptive weight shrinkage, and 5-seed averaging into a unified codebase.	4.370971	7.0118167688
41	Similar to 39, but blended fold-wise optimal weights using trimmed mean + shrink for conservative weighting, combined with 5-seed averaging.	—	—

Best Score: 🏆 7.0021252141 (Experiment 39, Public LB)

📄 Paper Study — Few-Shot Grounding DINO for Agriculture (Day 3)

Reviewed Experiments & Results, focusing on how few-shot embedding adaptation compares to zero-shot and fully fine-tuned baselines.

Datasets: Crop/weed detection, fruit counting, insect ID, wheat head detection, PhenoBench (agriculture); plus remote sensing datasets for validation.
Baselines: Zero-shot Grounding DINO, fully fine-tuned YOLO, and other SOTA models.
Key Results:
- Up to +24% mAP over YOLO on agriculture datasets.
- About +10% improvement over SOTA in remote sensing.
- Zero-shot works in simple scenes but fails in cluttered settings. Few-shot alleviates this.
Effect of Shot Number:
- 1-shot: sometimes worse than zero-shot (multi-class issues).
- 4-shot & 16-shot: steady performance gains.
- Instance segmentation: 8-shot showed strong improvements.
Takeaway: Few-shot adaptation offers large performance gains with minimal data, avoids prompt engineering, and generalizes well across domains.

💾 SQL Practice

Continued solving mock exam problems, focusing on query optimization and multi-table joins.
Maintaining steady progress toward the upcoming certification exam.

🎶 Personal Project — Bass Seeker

Began implementing bass_selector (v0.4), handling YouTube metadata retrieval and optional audio downloads.
Core features include:

Filtering candidates by duration, title keywords, and recency
Ranking by views, recency, and title cleanliness
Optional .wav conversion via ffmpeg
Logging and saving results to CSV for reproducibility

📄 Paper Study (Team) — Arm Wave Quality Assessment

Discussed potential approaches and novelty directions for an upcoming paper on evaluating the quality of arm wave movements.
Main points included:

How to collect and preprocess pose-based data for reliable evaluation
What baseline models to use for initial validation
Possible strategies for defining quality metrics and adapting them with learning frameworks
Considering a hybrid pipeline that balances classification and scoring aspects

The goal is to establish a clear research direction where the novelty can be well justified, while keeping the implementation both feasible and meaningful.

✅ TL;DR

📍 Fit 2 Give: Idea combining fitness goals with charitable giving
📍 Best public LB score 7.0021 with building-type–specific blending + seed averaging
📍 Reviewed Day 3 of few-shot adaptation paper (experiments & results)
📍 Practiced SQL mock problems for exam
📍 Developed bass_selector v0.4 for Bass Seeker
📍 Outlined novelty pipeline for arm wave quality assessment