Adaptive covariate acquisition for minimizing total cost of classification

Abstract In some applications, acquiring covariates comes at a cost which is not negligible. For example in the medical domain, order to classify whether patient has diabetes or not, measuring glucose tolerance can be expensive. Assuming that of each covariate, and misclassification specified by user, our goal minimize (expected) total classification, i.e. plus acquired covariates. We formalize this optimization using (conditional) Bayes risk describe optimal solution recursive procedure. Since procedure computationally infeasible, we consequently introduce two assumptions: (1) classifier represented generalized additive model, (2) sets are limited sequence increasing size. show under these assumptions, efficient exists. Furthermore, on several datasets, proposed method achieves most situations lowest costs when compared various previous methods. Finally, weaken requirement user specify all allowing minimally acceptable recall (target recall). Our experiments confirm target while minimizing false discovery rate covariate acquisition better than