A Clustering based Selection Framework for Cost Aware and Test-time Feature Elicitation

Abstract

Most learning algorithms are optimized with generalization and predictive performance as the goal. However, in most real-world machine learning applications, obtaining features at test time can incur a cost. For example, in clinical tasks, acquiring certain features such as FMRI or certain lab tests for patients can be expensive, while other features like patient demography or history are easily obtained and do not have a cost involved. Motivated by this, we address the problem of test-time elicitation of features. We formulate the problem of cost-aware feature elicitation as an optimization problem with trade-off between performance and feature acquisition cost. We assume that the cost of the features has already been paid in obtaining the training data. We propose a Clustering based Cost Aware Test-time Feature Elicitation (CATE) algorithm, which can select the relevant feature set given the observed attributes of the test instance. Our experiments on four real-world tasks demonstrate the efficacy and effectiveness of our proposed approach in both cost and performance.