Moving virtual boundary strategy for selective sampling

In relevance feedback of information retrieval system, selective sampling is often used to alleviate the burden of labeling by selecting only the most informative data to label. The traditional batch labeling model neglects the data's correlation and thus degrades the performance; while the theoretical optimal one-by-one training model is not efficient enough because of the high computational complexity. In this paper, we propose a Moving Virtual Boundary (MVB) strategy for informative data selection. We adopt a novel one-by-one labeling model, using the previous labeled data as extra guidance for the selection of next, and achieve better experimental results.