Economical Active Featurevalue Acquisition through Expected Utility Estimation.pdfVIP

To Appear in Proceedings of the KDD-05 Workshop on Utility-Based Data Mining, Chicago, IL, August 2005. Economical Active Feature-value Acquisition through Expected Utility Estimation Prem Melville Maytal Saar-Tsechansky Dept. of Computer Sciences Red McCombs School of Business Univ. of Texas at Austin Univ. of Texas at Austin melville@ maytal@ Foster Provost Raymond Mooney Stern School of Business Dept. of Computer Sciences New York University Univ. of Texas at Austin fprovost@ mooney@ ABSTRACT information that can be obtained at a low cost. In contrast, acquir- In many classification tasks training data have missing feature val- ing diagnostic test results from different health-care providers can ues that can be acquired at a cost. For building accurate predictive be significantly more expensive and time-consuming. Various solu- models, acquiring all missing values is often prohibitively expen- tions are available for learning models from incomplete data, such sive or unnecessary, while acquiring a random subset of feature as imputation methods [8], and learners that ignore missing feature values may not be most effective. The goal of active feature-value values such as the Naive Bayes classifier. However, these solu- acquisition is t