Active Data Acquisition with Incomplete Data
David P. Williams, Xuejun Liao, Lawrence Carin · 2005
We present a unified framework under which active data acquisition can be performed. This com-prehensive framework allows for the acquisition of both labels and features. Moreover, several types of feature acquisition are permitted, including the acquisition of individual or multiple features for individual or multiple data points, which may be either labeled or unlabeled. The algorithm chooses to acquire that data for which the expected benefit — defined as the cost of acquiring the data subtracted from the expected reduction in misclassification costs if the data is possessed — is a maximum. The algorithm automatically determines the most beneficial type of data to acquire when multiple options exist. The framework also has a natural, intuitive criterion for terminating the active data acquisition process: when the expected ben-efit of all possible acquisitions is no longer positive. We also present a classifier that utilizes incomplete data, which is then employed in the proposed active data acquisition framework. Experimental results demonstrate the superiority of the proposed approach over random data acquisition. I.