π§ Daily Study Log [2025-08-14]
Focused on advanced model blending techniques in the electricity forecasting competition, documented a new urban sound-related idea, and continued paper review and SQL practice.
Achieved another best public LB score with improved stability through repeated folds and seed averaging.
Proposed a concept for mapping and visualizing urban noise patterns by integrating sound level meters, citizen reports, and environmental data.
The goal is to identify high-noise areas, analyze noise sources, and support urban planning decisions to improve quality of life.
π View idea
Continued building-typeβspecific modeling with enhanced blending strategies:
| No. | Description | Local SMAPE | Public LB SMAPE |
|---|---|---|---|
| 36 | Split data by building type, tuned XGBoost and LightGBM via RandomizedSearchCV, applied 7-Fold OOF CV, and blended predictions with optimal weights based on OOF results. (best) | 3.982886 | 7.1161849075 |
| 37 | Extended 36 by using RepeatedKFold to reduce variance, refined OOF-based weight search to 0.01 increments, and retrained with multiple seeds for seed averaging to mitigate overfitting. | 3.982668 | 7.0759340681 |
| 38 | Computed fold-wise optimal weights, blended using the average of these weights, and applied 5-seed averaging for further variance reduction. Preprocessing remained unchanged. | β | β |
Best Score: π 7.0759340681 (Experiment 37, Public LB)
Reviewed methods for improving object detection in agricultural contexts using embedding-based adaptation for few-shot scenarios.
Focus was on how embedding space modifications enable robust classification with minimal labeled data.
π View notes
Solved past exam problems, focusing on query optimization and multi-table joins.
π Urban Noise Palette: Concept for urban noise analysis & visualization
π Best public LB score with RepeatedKFold + seed averaging blending
π Continued few-shot detection paper review
π Practiced SQL exam problems