Method for Generating Explainable Deep Learning Models in the Context of Air Traffic Management

• Published in: Machine Learning, Optimization, and Data Science Vol. 13163; pp. 214 - 234
• Main Authors: Rudd, Keith, Eshow, Michelle, Gibbs, Mary
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Air traffic management; Deep learning; Explainable AI
• ISBN: 9783030954666; 3030954668
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-95467-3_17

Model explainability, interpretability, and explainable AI have become major research topics, particularly for deep neural networks where it is unclear what features the network may have used to come to a particular output. This paper presents a method for understanding these features. The method processes auto-encoder outputs to generate feature vectors and allows a user to explore these features and gain insight into the network’s behavior. The method is applied to the U.S. air traffic management domain, an area rich in data and complexity, where use of deep learning is growing rapidly. In particular, the method is used to create models to predict throughput at major airports based on current and predicted weather and air traffic conditions. Models are created for two major U.S. airports of interest to NASA. For this application, the method is found to produce acceptable levels of prediction accuracy. It also produces explanations that relate input conditions to each other and to the predictions in ways that align with those of subject matter experts. Acceptance by subject matter experts is necessary if deep learning models are to be adopted in the air traffic management domain.