We study on-the-fly classification of evolving text streams in which the relation between the input data target labels changes over time—i.e. “concept drift”. These variations decrease the model’s performance, as predictions ...