Linear adaptive filtering for regression in data streams

Many applications in supervised learning of evolving data streams deal with regression, since they need to forecast numeric values in a fast and accurate way. Most of the current regression methods are inspired by the state-of-the-art data stream classifiers adapted to the regression problem. In this paper, we present a novel approach to the problem based on classical linear adaptive filtering theory. The linearity allows to obtain simple general models highly accurate without needing to use ensemble learners. We study the recursive version of these methods as they are the ones that satisfy the stream requirements and we include in their formulation the ability to track the drift of the data in stream scenarios. We show how they integrate in an easy and fast way the adaptation to changing environments avoiding the use of explicit drift detectors. We apply these methods to classical datasets in stream regression and show how they can outperform them in drifting cases when the linear model is a good approximation to the problem.