Feature-Based Priority Queuing

Traditional queuing theory assumes types are known or perfectly observed, and each type is typically put in its type-specific queue which is prioritized using some version of the celebrated c-mu rule; we call this type-based queueing. We study feature-based priority queuing where types are not perfectly observed but are inferred from observed features using a ``classifier.'' A practically appealing approach combines an off-the-shelf classifier that predicts the type with type-based priority queueing. We propose a direct approach that optimizes the classifier to directly predict the priority queue from features. The explicit modeling of the classifier in the queueing-system design is the novel contribution of this paper. We present an analytic model to study the optimal queue classification that minimizes queuing delay costs. We study how the optimal number of priority queues and the assignment of features to queues changes with the classifier accuracy. We present a numerical study on a real data set of medical images utilized in digital triage in radiology. We find that, relative to type classification, optimal feature-based priority queuing can improve delay costs by up to 54% using state-of-the-art image classifiers