Theoretical and empirical analysis of diversity in non-stationary learning

In non-stationary learning, we require a predictive model to learn over time, adapting to changes in the concept if necessary. A major concern in any algorithm for non-stationary learning is its rate of adaptation to new concepts. When tackling such problems with ensembles, the concept of diversity appears to be of significance. In this paper, we discuss how we expect diversity to impact the rate of adaptation in non-stationary ensemble learning. We then analyse the relation between voting margins and a popular measure of diversity, KW variance, and use the similarities between them to draw some useful conclusions regarding ensemble adaptivity.