Dealing with Change

Abstract

A central feature of the data stream model is that streams evolve over time, and algorithms must react to the change. For example, let us consider email spam classifiers, which decide whether new incoming emails are or are not spam. As classifiers learn to improve their accuracy, spammers are going to modify their strategies to build spam messages, trying to fool the classifiers into believing they are not spam. Customer behavior prediction is another example: customers, actual or potential, change their preferences as prices rise or fall, as new products appear and others fall out of fashion, or simply as the time of the year changes. The predictors in these and other situations need to be adapted, revised, or replaced as time passes if they are to maintain reasonable accuracy. In this chapter we will first discuss the notion of change as usually understood in data streaming research, remarking on the differences with other areas that also consider changing data, then discuss possible types of change, some measures of performance used in the presence of change, and the general strategies for designing change-aware algorithms on streams (section 5.1). We will then consider methods for accurately estimating statistics of evolving data streams (section 5.2), and methods for detecting and quantifying change (section 5.3). We will conclude with some references on time-aware sketches and change detection in multidimensional streams (section 5.4). There is a vast literature on change detection in data, and we have prioritized methods that are most applicable to data streams, easy to implement, and computationally light, and in particular those that are available in MOA. A comprehensive discussion for all types of data is [27], and an up-to-date discussion for streams is [116]. A framework for experimenting with change detection and management in MOA is presented in [41].