Smart Warehouse Delivery Delay Prediction AI Competition

최대한 이것저것 넣고싶어서 하다보니 모델이랑 구조가 복잡해서 합치는데도 시간이 오래걸렸네요,,

[Raw train/test/layout]
        ↓
[Feature Engineering]
        ↓
 ┌─────────────────────┐
 │ Tabular Models       │
 │ LGBM / HistGB / Cat  │
 └─────────────────────┘
        ↓
 [OOF / Test predictions]
        ↓
 ┌─────────────────────┐
 │ Sequence Models      │
 │ GRU / CNN / TCN      │
 └─────────────────────┘
        ↓
 [Tabular + Sequence Blend]
        ↓
 [Residual / Meta / Gate / Calibration]
        ↓
 [Late traffic / pressure / nonnegative correction]
        ↓
 [Final submission]

의 흐름으로 학습을 진행했고 train/test/layout에서 시간·layout·traffic feature를 만든 뒤, LGBM/HistGB/CatBoost 같은 tabular 모델과 GRU/CNN/TCN 같은 sequence 모델을 각각 학습합니다.

그 예측값들을 OOF 기준으로 blend하고, residual/meta/gate/traffic 보정까지 쌓아서 최종 v509 submission을 만드는 ensemble pipeline입니다.