🧠 Daily Study Log [2025-08-19]
Generated a new idea, practiced SQL problems intensively for the upcoming exam, conducted multiple competition experiments with blending strategies, wrapped up Day 4 of the few-shot adaptation paper with conclusions and future work, created a small adapter implementation, and archived the Bass Seeker project with reflections on coding practices.

💡 Idea Generation — Pink Sky Predictor

Proposed an idea that explores predicting the occurrence of pink skies using weather and environmental data.
This project concept combines atmospheric science with aesthetic prediction.

🔗 View idea

💾 SQL Practice

Continued preparing for the certification exam on August 23, focusing on solving past exam problems.
Emphasis on query optimization and handling multi-table join scenarios.

🏆 Competition — Electricity Forecasting

Conducted several experiments refining building-type–specific models and blending methods:

No.	Description	Local SMAPE	Public LB SMAPE
39	Best score: tuned XGB/LGBM via RS-CV for each building type, computed fold-wise optimal weights on 7-Fold OOF, averaged them, and 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	Used RS-CV per building type, blended fold-wise weights with trimmed mean + shrink, and applied 5-seed averaging.	12.496758	6.9964770263
42	Partitioned by building type, tuned with lightweight RS-CV (n_iter=6), blended with trimmed mean + shrink, applied 3-seed averaging and post-processing (0/top clipping, building-level median filter, OOF bias correction).	12.553428	15.1738844787
43	Single-model XGB with RS-CV (n_iter=8), TimeSeriesSplit(7) OOF validation, 5-seed averaging, and building-level median filtering. Disabled bias/top clipping for leaderboard safety.	16.282731	7.1981428039

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

📄 Paper Study — Few-Shot Grounding DINO for Agriculture (Day 4)

Reviewed Conclusion, Limitations, and Future Work of the paper.

Conclusion: Proposed a few-shot embedding adaptation method, replacing text prompts with learnable embeddings, achieving strong improvements over zero-shot and full fine-tuning baselines.
Limitations: 1-shot unstable, challenges in complex tasks, sensitive to hyperparameters.
Future Work: Cross-domain extension (e.g., medical imaging), hierarchical embeddings, temporal/segmentation tasks, and hybrid fine-tuning.

Additionally, implemented a small adapter module to explore the idea hands-on.

🎶 Personal Project — Bass Seeker (Archived)

Archived the Bass Seeker project.
🔗 Archive Link

Reflection after participating in the Strawberry competition:

Realized gaps in code understanding and clarity.
Moving forward, I plan to write code more carefully and thoroughly understand every component.

✅ TL;DR

📍 Proposed Pink Sky Predictor idea
📍 SQL exam prep with past problems (exam on Aug 23)
📍 Best LB score 7.0021 from building-type blending + seed averaging
📍 Finished Day 4 of few-shot adaptation paper & built an adapter prototype
📍 Archived Bass Seeker, resolved to improve coding practices