Smart Warehouse Delivery Delay Prediction AI Competition

최종 성적: Public LB 9.802014278 / Private 10.01488 / 18등

PB-style 피처 엔지니어링
로봇 운영 지표의 lag/rolling/비율 피처 714개에 시나리오 내 25개 타임스텝 집계(mean/std/max/min) 84개를 추가, 총 798개 피처를 구성했습니다. 
시나리오 집계 피처가 GBDT blend CV 8.54 → 8.48 개선의 핵심이었습니다. 

이종 모델 9종
LightGBM(MAE·Huber), XGBoost(MAE), CatBoost(MAE), TabNet, Transformer(Pre-Norm d=256 L=6 L1Loss), MLP(TabMLP L1Loss)를 3-seed × 5-fold GroupKFold로 학습하여 다양성을 확보했습니다.

불확실성 메타 스태킹
9개 OOF 예측값 외에 예측 std와 range를 메타 피처로 추가한 LGB 스태킹(num_leaves=15) 적용. 불확실성 피처 추가가 Public LB 9.821 → 9.802 개선의 핵심 기여입니다.

GroupKFold
scenario_id 기준 GroupKFold(5)로 시나리오 단위 정보 누수를 방지. GBDT는 log1p 타깃 변환, Transformer/MLP는 L1Loss 직접 학습.

재현
Score 복원: run_experiments_v8.py → run_v8_transformer.py --v4 → run_v8_mlp.py → run_stacking.py --tag v9
폴더: ./data/(데이터), ./checkpoints_v8/(GBDT), ./checkpoints_v8_transformer/(DL), ./checkpoints_v8_tabnet/(TabNet), ./checkpoints_strict/(strict2_tabnet)

체크포인트는 깃허브에 있습니다.
github: https://github.com/kgh2895/dacon-smart-warehouse-2026.git