Novel class detection in data streams

Abstract

Data stream classification is challenging process as it involves consideration of many practical aspects associated with efficient processing and temporal behavior of the stream. Two such aspects which are well studied and addressed by many present data stream classification techniques are infinite length and concept drift. Another very important characteristic of data streams, namely, conceptevolution is rarely being addressed in literature. Concept-evolution occurs as a result of new classes evolving in the stream. Handling concept evolution involves detecting novel classes and training themodel with the same. It is a significant technique to mine the data where an important class is under-represented in the training set. This paper is an attempt to study and discuss the technique to handle this issue. We implement one of such state-of-art techniques and also modify for better performance.