Exploring Early Classification Strategies of Streaming Data with Delayed Attributes

In contrast to traditional machine learning algorithms, where all data are available in batch mode, the new paradigm of streaming data poses additional difficulties, since data samples arrive in a sequence and many hard decisions have to be made on-line. The problem addressed here consists of classifying streaming data which not only are unlabeled, but also have a number l of attributes arriving after some time delay τ . In this context, the main issues are what to do when the unlabeled incomplete samples and, later on, their missing attributes arrive; when and how to classify these incoming samples; and when and how to update the training set. Three different strategies (for l = 1 and constant τ ) are explored and evaluated in terms of the accumulated classification error. The results reveal that the proposed on-line strategies, despite their simplicity, may outperform classifiers using only the original, labeled-and-complete samples as a fixed training set. In other words, learning is possible by properly tapping into the unlabeled, incomplete samples, and their delayed attributes. The many research issues identified include a better understanding of the link between the inherent properties of the data set and the design of the most suitable on-line classification strategy.