Active Learning: Applications, Foundations and Emerging Trends

Active learning optimizes the interaction between artificial data mining/machine learning systems and humans. For example, its techniques are used for selecting the most relevant information to be requested, or for determining the most informative experiment to be performed. With increasing volumes of data, contrasting limited human supervision capacities, the optimization of this interaction has become even more important. Hence, active sampling and data acquisition techniques could contribute to the design and modeling of highly intelligent learning systems. This tutorial presents the basic techniques for pool-based and on-line active learning for streams. It contains a summary of the common concepts like version space partitioning, uncertainty sampling and decision theoretic approaches, and shortly mentions the connection between reinforcement learning and active learning. We show how these concepts can be used in data streams and on-line applications, and discuss the main challenges of stream active learning. Finally, we evaluate frameworks for pool-based and stream-based active learning to validate if a method is applicable for a specific demand