[Paper Review] AI-CARGO: A Data-Driven Air-Cargo Revenue Management System

Stefano Rizzo, Ji Lucas+3 more|arXiv (Cornell University)|May 22, 2019

Data Quality and ManagementDecision Sciences20 references3 citations

TL;DR

AI-CARGO is a data-driven air-cargo revenue management system that integrates machine learning-based volume prediction with optimization to reduce offloading costs and improve revenue. By detecting disguised missing values (DMVs) in booking data and using Gradient Boosting Machines (GBMs) to predict actual cargo volume, it enables dynamic acceptance/rejection decisions, reducing revenue loss from booking discrepancies in volume-constrained flights.

ABSTRACT

We propose AI-CARGO, a revenue management system for air-cargo that combines machine learning prediction with decision-making using mathematical optimization methods. AI-CARGO addresses a problem that is unique to the air-cargo business, namely the wide discrepancy between the quantity (weight or volume) that a shipper will book and the actual received amount at departure time by the airline. The discrepancy results in sub-optimal and inefficient behavior by both the shipper and the airline resulting in the overall loss of potential revenue for the airline. AI-CARGO also includes a data cleaning component to deal with the heterogeneous forms in which booking data is transmitted to the airline cargo system. AI-CARGO is deployed in the production environment of a large commercial airline company. We have validated the benefits of AI-CARGO using real and synthetic datasets. Especially, we have carried out simulations using dynamic programming techniques to elicit the impact on offloading costs and revenue generation of our proposed system. Our results suggest that combining prediction within a decision-making framework can help dramatically to reduce offloading costs and optimize revenue generation.

Motivation & Objective

• Address the critical problem of volume discrepancy between booked and actual cargo shipments in air-cargo, which leads to inefficiencies and revenue loss.
• Improve revenue management in air-cargo by integrating prediction and decision-making, unlike traditional systems focused only on passenger revenue.
• Develop a robust data cleaning pipeline to handle heterogeneous, low-quality booking data, particularly disguised missing values (DMVs).
• Create a prediction-driven optimization framework that models revenue decisions based on predicted actual cargo volume at departure.
• Deploy and validate the system in a real airline production environment to demonstrate tangible improvements in offloading cost and revenue generation.

Proposed method

• Introduce a data cleaning module that detects and resolves disguised missing values (DMVs) in booking data, improving input quality for prediction.
• Train a Gradient Boosting Machine (GBM) model to predict the actual received volume (rcsvol) based on booking features, with special handling for DMV-affected entries.
• Formulate the air-cargo revenue management problem as a sequential, stochastic optimization task where the state (total volume) is realized at departure.
• Integrate the GBM prediction model into a decision-making framework that evaluates whether to accept or reject a booking based on predicted volume and revenue impact.
• Use dynamic programming simulations to evaluate the system’s performance under varying booking variance and capacity constraints.
• Compare the proposed D1S policy against FCFS (First-Come, First-Served) using synthetic and real datasets to assess revenue and offloading cost performance.

Experimental results

Research questions

• RQ1 How does the presence of disguised missing values (DMVs) in air-cargo booking data affect the accuracy of volume prediction models?
• RQ2 To what extent can a machine learning-based volume prediction model reduce the discrepancy between booked and actual cargo volume in air-cargo operations?
• RQ3 How does integrating prediction with optimization improve revenue and reduce offloading costs compared to traditional FCFS policies?
• RQ4 What is the impact of varying booking variance (θ) on the performance of the D1S decision policy versus FCFS in simulated air-cargo scenarios?
• RQ5 Can a prediction-driven optimization framework effectively handle the dynamic and stochastic nature of air-cargo capacity utilization?

Key findings

• The detection and resolution of disguised missing values (DMVs) significantly improved the quality of volume predictions, directly enhancing downstream decision-making.
• The GBM-based prediction model reduced bias and variance in volume forecasts, especially at the flight level, leading to more reliable decision inputs.
• In dynamic programming simulations, the D1S policy outperformed FCFS in both expected revenue and final revenue after offloading costs, with a consistent advantage across all tested variance levels (θ = 0.8 and θ = 1.0).
• At θ = 1.0, D1S achieved an expected revenue of 2927.87 and final revenue of 2553.74, compared to FCFS’s 2548.95 and 2517.23, respectively, demonstrating a sustained revenue benefit even after offloading costs.
• Table 5 shows that D1S dynamically accepts high-revenue-rate shipments even after capacity is reached, while FCFS rejects them, leading to higher revenue accumulation and better handling of volume uncertainty.
• The system successfully reduced offloading costs and improved revenue generation by proactively managing bookings based on predicted actual volume, validating the value of integrating prediction with optimization.