Learning from streaming data represents an important and challenging task. Maintaining an accurate model, while the stream goes by, requires a smart way for tracking data changes through time, originating concept drift. One way to treat this kind of problem is to resort to ensemble-based techniques. In this context, the advent of new technologies related to web and ubiquitous services call for the need of new learning approaches able to deal with structured-complex information, such as trees. Kernel methods enable the modeling of structured data in learning algorithms, however they are computationally demanding. The contribute of this work is to show how an effective ensemble-based approach can be deviced for streams of trees by optimizing the kernel-based model representation. Both efficacy and efficiency of the proposed approach are assessed for different models by using data sets exhibiting different levels and types of concept drift.