Machine Learning-Based Optimization of Product Loading in Airline In-Flight Sales
Bildiriler > Tam Metin Bildiri
Yazarlar: Serkan Kocoglu, Züleyha AKUSTA DAĞDEVİREN
2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA), , , Uluslararası, Tam Metin Bildiri, ss 1-10
DOI: 10.1109/ichora69329.2026.11537189
Özet
Airline in-flight retail operations face a critical inventory challenge: determining the optimal quantity of perishable products to load onto each flight. Overstocking leads to spoilage, disposal, and unnecessary fuel costs from transporting excess weight, while understocking results in lost revenue and diminished passenger satisfaction. Traditional forecasting models are typically trained to minimize symmetric error metrics, which fail to reflect the asymmetric economic consequences of overstocking and understocking. This study proposes a cost-aware machine learning framework that reformulates this inventory challenge as an asymmetric optimization task grounded in the classical Newsvendor model. A custom loss function is introduced to explicitly model the economic penalties associated with these decisions. The loss function is integrated into a multi-strategy ensemble framework combining Random Forest, LightGBM, CatBoost, and a Zero-Inflated LightGBM model. The ensemble employs four optimization strategies: multi-start global weight optimization via Differential Evolution, Newsvendoraligned quantile-shrinkage calibration, per-level joint weight-bias optimization across minimum-stock constraints, and cost-guided integer rounding. The framework is evaluated on 362,732 empirical flight records drawn from a commercial airline. Results show that the proposed approach reduces total operational cost to $679,680, corresponding to a 55.4 % reduction compared to the human-decision baseline of $1,525,355, while strictly adhering to minimum service constraints. The findings demonstrate that aligning machine learning objectives with operational cost structures can substantially improve decision quality in revenue-critical retail contexts with asymmetric cost structures.