Dynamic Clustering Forest: An ensemble framework to efficiently classify textual data stream with concept drift

Textual stream mining with the presence of concept drift is a very challenging research problem. Under a realistic textual stream environment, it often involves a large number of instances characterized by a high-dimensional feature space. Accordingly, it is computationally complex to detect concept drift. In this paper, we present a novel ensemble model named, Dynamic Clustering Forest (DCF), for textual stream classification with the presence of concept drift. The proposed DCF ensemble model is constructed based on a number of Clustering Trees (CTs). In particular, the DCF model is underpinned by two novel strategies: (1) an adaptive ensemble strategy to dynamically choose the discriminative CTs according to the inherent property of a data stream, (2) a dual voting strategy that takes into account both credibility and accuracy of a classifier. Based on the standard measure of Mean Square Error (MSE), our theoretical analysis demonstrates the merits of the proposed DCF model. Moreover, based on five synthetic textual streams and three real-world textual streams, the results of our empirical tests confirm that the proposed DCF model outperforms other state-of-the-art classification methods in most of the high-dimensional textual streams.