Prediction-time Active Feature-value Acquisition for Cost-effective Customer Targeting

In general, the prediction capability of classification models can be enhanced by acquiring additional relevant features for instances. However, in many cases, there is a significant cost associated with this additional information — driving the need for an intelligent acquisition strategy. Motivated by real-world customer targeting domains, we consider the setting where a fixed set of additional features can be acquired for a subset of the instances at test time. We study different acquisition strategies of selecting instances for which to acquire more information, so as to obtain the most improvement in prediction performance per unit cost. We apply our methods to various targeting datasets and show that we can achieve a better prediction performance by actively acquiring features for only a small subset of instances, compared to a random-sampling baseline.