Cooperative Agents for Discovering Pareto-Optimal Classifiers Under Dynamic Costs

Abstract

In contrast to passive classifiers that use all available input feature values to assign class labels to instances, active classifiers determine the features on which to base the classification. Motivated by the tradeoff between the cost of classification errors and the cost of obtaining additional information, active classifiers are widely used for diagnostic applications in domains such as in medicine, engineering, finance, and natural language processing. This paper extends the extant literature on active classifiers to applications where cost of obtaining additional information may vary over instances to be classified and over time. We show that this entails training a set of classifiers that grows exponentially with the number of features and propose an efficient way to discover models in the cost-accuracy Pareto optimal frontier. Our method is based on a set of cooperative agents. The incremental contributions of agents to a coalition is used as a surrogate measure to guide a heuristic search for models. Empirical results based on controlled experiments indicate that our approach can identify Pareto-optimal active classifiers under dynamic costs even in domains that involve a large number of input features.