Probabilistic Active Learning in Datastreams

In recent years, stream-based active learning has become an intensively investigated research topic. In this work, we propose a new algorithm for stream-based active learning that decides immediately whether to acquire a label (selective sampling). To this purpose, we extend our pool-based Probabilistic Active Learning framework into a framework for streams. In particular, we complement the notion of usefulness within a topological space (“spatial usefulness”) with the concept of “temporal usefulness”. To actively select the instances, for which labels must be acquired, we introduce the Balanced Incremental Quantile Filter (BIQF), an algorithm that assesses the usefulness of instances in a sliding window, ensuring that the predefined budget restrictions will be met within a given tolerance window. We compare our approach to other active learning approaches for streams and show the competitiveness of our method.