Survey on Categorization and Detection of Adaptive Novel Class of Feature Evolving Data Streams

Classification in the data stream is the challenging fact for the data mining community. In this paper, we tackle four major challenges which are infinite length, concept drift, concept evolution, and feature evolution. As we know that the data streams are huge in amount, so practically it is not possible to store the data and used it for the training purpose. The results of changes in the underlying concepts are occurred because of concept drift, which is the general observable fact in the data streams. The result of new classes surfacing in the data streams occurs because of concept evaluation. The feature evaluation generally occurs in many streams like text streams, in text streams new features emerge as stream advancement. Many existing methods of the data stream classification tackle only first two challenges and ignore last two challenges. Here in this paper we proposed an ensemble classification skeleton, in which each classifier is prepared with a novel class detector to tackle the concept drift and concept evolution. We also proposed the feature set homogenization methods for feature evaluation. We improve the component of novel class detection by making it more adaptive to the evolving stream, and enable it to notice more than one novel class at a time. As comparing with the existing methods of the novel class detector method the efficiency of the proposed method is more than the existing one.