Efficient computation of queries on feature streams

Abstract

This article introduces the notion of virtual feature stream , a feature stream defined from a primary data stream, in which at any time only the features that are needed to compute the queries that are currently running in the system are computed. Virtual feature streams are, in general, impossible to determine a priori, but the paper introduces an algorithm that stops the computation of features as soon as it can be proved that they are no longer needed thus generating, albeit in a roundabout and more expensive than the ideal way, a feature stream that is less expensive than the complete one to compute and safe: the queries that accept the virtual feature stream are those (and only those) that would accept the original feature stream.