Economical Active Feature-value Acquisition through Expected Utility Estimation

In many classification tasks training data have missing feature valuesthat can be acquired at a cost. For building accurate predictive models,acquiring all missing values is often prohibitively expensive orunnecessary, while acquiring a random subset of feature values may notbe most effective. The goal of active feature-value acquisition is toincrementally select feature values that are most cost-effective forimproving the model’s accuracy. We present two policies, SampledExpected Utility and Expected Utility-ES, that acquire feature valuesfor inducing a classification model based on an estimation of theexpected improvement in model accuracy per unit cost. A comparison ofthe two policies to each other and to alternative policies demonstratethat Sampled Expected Utility is preferable as it effectively reducesthe cost of producing a model of a desired accuracy and exhibits aconsistent performance across domains.