A0514
Title: Iterative modeling under feature acquisition constraints
Authors: Marshall Honaker - University of Pittsburgh (United States) [presenting]
Lucas Mentch - University of Pittsburgh (United States)
Meredith Wallace - University of Pittsburgh (United States)
Abstract: The traditional approach to statistical modeling involves collecting a set of relevant features, performing some exploratory data analysis, and fitting a model to the response. Once fit, the model is used to evaluate new observations and obtain predictions. An implicit but often overlooked assumption in this setup is that all features must be collected for every new observation. In practice, the real-world cost or burden of acquiring a feature's value must be balanced with its utility as a predictor. For example, medical professionals may employ different tests in different situations based on their availability, affordability, and invasiveness. A novel perspective on cost-sensitive learning is proposed that sequentially introduces features in order of increasing cost. Rather than a single model that balances accuracy and feature cost, a sequence of models is produced that optimizes performance subject to the allotted cost at each stage. Moreover, each model is equipped with a reject option so that predictions are made when model confidence is sufficiently high and deferred to the next most costly model otherwise. In medical contexts, this allows patients to be diagnosed using the least costly measures without sacrificing predictive accuracy. Numerous demonstrations of this model-agnostic framework are provided.
