🧠 Daily Study Log [2025-08-03]
From creative computer vision ideas to leaderboard-breaking ensemble tuning — today was a good balance of ideation and implementation.

💡 Project Idea — Taste by Color: Visual Food Flavor Predictor

Proposed a fun computer vision system that predicts flavor profiles (sweet, spicy, sour, etc.) based on food image color features.

Inspired by human heuristics like:

“Red-orange = spicy?” “Light brown = sweet?”

Core concept:

• Use dominant HSV color features
• Train a classifier to predict taste labels
• Possibly use recipes or crowdsourced labels for supervision

🔗 View full idea

🧪 Competition — Electricity Usage Forecasting (DACON)

22. Time × Building Type Voting Ensemble

Split data by:

• Time: Night / Day / Evening
• Building Type: Top 7 + Others

For each of the resulting 24 subgroups, trained a VotingRegressor (XGB + LGBM + GBR).
Also introduced new feature combinations:

• hour_sin, hour_cos (cyclical time features)
• 냉방_x_시간, 면적_x_요일 (crossed variables)

📊 Results

• Local SMAPE: 11.7492
• Public LB: 9.61451 ✅ Best score so far

✅ TL;DR

📍 Idea: Taste-by-Color — food image → flavor profile classifier
📍 Competition: Best-performing Voting Ensemble using Time × Building segmentation (SMAPE: 9.6145)