Keynote speaker 1: Active online learning

Abstract

Summary form only given. Over the recent years learning from data streams that evolve over time has been witnessing an ever-increasing interest within research and industry communities. Typically a wide range of applications exploit data streams for different sorts of decision making, including monitoring, industrial processes, internet traffic, surveillance, etc. By their very nature, data streams are usually unlabeled given the high velocity of their generation. Collecting labelled examples become very difficult, delayed, costly and sometimes prone to errors. It is therefore very important to devise mechanisms to optimize the labeling process. Active learning offers a principled and systematic way to selectively choose candidate data examples whose labels are to be queried. The overall goal of active learning is to provide, in the worst case, the same performance as that of passive learning (i.e., relying on random sampling) while using less labeled examples. Obviously, the learner should also be able to accommodate unlabeled and labeled data in an online manner. In this talk we will cover recent work on active learning for data stream classification, which is known as stream-based selective sampling. In this latter, the learner makes immediate query decision for each data example during a single scan of the data stream. Stream-based selective sampling is in particular suitable for applications that demand on-the-fly interactive labelling. It is however difficult, because the learner lacks complete knowledge of the underlying data distribution and because such distribution may suffer dynamic change over time. We will overview active learning for stationary as well as non-stationary evolving data streams. In particular, we will discuss multi-criteria active learning and methods for dealing with data drift using online active learning. We will also highlight some of the typical applications where online active learning is relevant.