Deep Learning in Partially-labeled Data Streams

Of the considerable research on data streams, relatively little deals with classification where only some of the instances in the stream are labeled. Most state-of-the-art data-stream algorithms do not have an effective way of dealing with unlabeled instances from the same domain. In this paper we explore deep learning techniques that provide important advantages such as the ability to learn incrementally in constant memory, and from unlabeled examples. We develop two deep learning methods and explore empirically via a series of empirical evaluations the application to several data streams scenarios based on real data. We find that our methods can offer competitive accuracy as compared with existing popular data-stream learners.