Active Feature-Value Acquisition

摘要: Most induction algorithms for building predictive models take as input training data in the form of feature vectors. Acquiring values features may be costly, and simply acquiring all wasteful or prohibitively expensive. Active feature-value acquisition (AFA) selects incrementally an attempt to improve model most cost-effectively. This paper presents a framework AFA based on estimating information value. Although straightforward principle, estimations approximations must made apply practice. We present policy, sampled expected utility (SEU), that employs particular enable effective ranking potential acquisitions settings where relatively little is available about underlying domain. then experimental results showing that, compared with policy using representative sampling acquisition, SEU reduces cost producing desired accuracy exhibits consistent performance across domains. also extend more general modeling setting which well class labels are missing costly acquire.