Machine Learning in Java - Second Edition

Active learning

We all know that classifiers work well with labeled data, but that is not always the case with stream data. For example, the data from a stream may come unlabeled. Labeling data is costly, because it requires human intervention to label the unlabeled data. We understand that the streams generate large amounts of data. Active learning algorithms only do the labeling for selective data. The data to be labeled is decided on from historical data suited for pool-based settings. Regular retraining is required to decide whether a label is required for incoming instances. A simple strategy for labeling data is to use a random strategy. It is also called a baseline strategy, and it asks for a label for each incoming instance with probability ...