The ability to accurately forecast the amount of passengers that will board a particular flight is crucial for airline operations. But how do we design a machine learning algorithm for this use case and in what ways can we improve it? In this talk, we start with a simple linear model, evolving to increasingly complex deep learning neural network architectures.
Predictions for the number of airline passengers that will board a particular flight have a variety of uses, including revenue management, fuel planning, the anticipation of no-shows or excess hand luggage, and catering supply chain management. In this talk, we outline the business requirements for a passenger forecast system, formalize the machine learning problem at hand, discuss ways to evaluate model candidates, and subsequently advance through a series of tailor-made algorithms with increasing complexity. From shortcomings of each model candidate, we introduce a new modular improvement to the algorithm. We build from a simple linear model to more sophisticated deep learning neural network architectures, including advanced methods such as conditional density estimation and sequence embeddings. We intersperse conceptual model design with snippets of Python code to provide practical handles. In conclusion, we briefly delineate some open modeling challenges associated with this use case.