Calibrated ML and Decision-Curve Analysis for Passenger-Service Operations

Overview This repository accompanies the manuscript **“From Probabilities to Workload: Calibrated ML and Decision-Curve Analysis for Passenger-Service Operations”** (submitted to the *International Journal of Transportation Science and Technology*). It provides a **fully reproducible pipeline** for training, calibrating, and evaluating machine‑learning models for airline passenger‑service operations, with a direct translation from **calibrated risk** to **operational workload** (alerts/day, TP/FP/day) using **Decision‑Curve Analysis (DCA)** within an operational policy window. Contents - `code/` — Python source (data pipeline, model training, isotonic calibration, metrics: Brier/Murphy/ECE, DCA, explainability). - `configs/` — YAML configuration and schema files (frozen preprocessing; leakage‑controlled splits). - `artifacts/` — Derived artifacts: model binaries, predictions, logs, figures, and tables referenced in the manuscript. - `runbook.md` — Minimal commands to reproduce the main results end‑to‑end. - `environment.txt` — Pinned Python versions / key dependencies for reproducibility. - `LICENSE` — Open‑source license for the code (MIT recommended). - `CITATION.cff` — How to cite this repository. - `README.md` — Quick start and usage notes. Data: This deposit includes *derived* artifacts only. The raw passenger dataset is publicly available from the original provider and should be obtained directly by users following the instructions in `README.md`. All paths/commands assume the raw data are placed under `data/raw/` as described in the runbook. Methods snapshot - Models: regularized logistic regression, random forests, and gradient‑boosting trees; probabilities calibrated with **isotonic regression**. - Reliability: **Brier score** with **Murphy decomposition** (Reliability/Resolution/Uncertainty) and **Expected Calibration Error (ECE)**. - Decision impact: **Net benefit / Decision‑Curve Analysis** in a policy window; translation to workload per operational segment (cabin class, distance, delay band, travel type).