rhlprj/container-appt-ranker

Container Appointment Ranking

Calibrated dual-head LightGBM model for ranking container appointment slots.

Architecture: GBDT + Neural Blend (Option 3)

• LightGBM for tabular features + hash-bucket embeddings
• Dual-head: Head A (P_accept) + Head B (P_zero_rehandle)
• Isotonic calibration on realized val slots
• Combined score = 0.6 * P_accept + 0.4 * P_zero_rehandle

Metrics

	Baseline	Proposed	Target
AUC Head A	0.638	0.710	≥ 0.82
Brier Head A	0.186	0.174	≤ 0.18
AUC Head B	0.724	0.723	≥ 0.74
NDCG@3	0.726	0.922	≥ 0.68
Cold AUC	0.639	0.688	≥ 0.72

Files

• inference.py: predict(container_features, trucker_id, candidate_slots)
• data/synth.py: Synthetic data generator (seed=42)
• train.py / train_slots.py: Baseline + proposed training
• eval.ipynb: Reproduces all metrics
• model_card.md: Intended use, limitations, failure modes