Novel Class Detection for Feature Evolving Data Streams

Abstract

Data Stream Mining is the method of deriving knowledge from constant and quickly developing records of information. A data stream is an ordered grouping of events. These occurrences can be read just once or a less number of times utilizing restricted using limited storage capabilities and computing.Examples of such data streams include ATM exchanges, sensor information, telephone discussions and so forth. Data stream characterization has many difficulties in the information mining field. In this paper, we manage four noteworthy difficulties in the field of Data stream characterization which are infinite length, concept-drift, concept-evolution and feature-evolution. A data stream is never-ending in length, hence it is not practical to store and utilize all the historical data for training purpose. Concept-drift occurs as a result of changes in the fundamental concepts. Concept-evolution happens when new classes develop in the information data. An example of concept-evolution is Twitter, where new themes develop routinely in the stream of instant messages.Feature-evolution is a regularly happening process in data streams, where new features evolve and old features vanish. We investigate this problem during this paper, and propose improved solutions. Our current work additionally addresses the feature-evolution problem in data streams, like text streams, wherever new options (words) emerge and previous options disappear.We moreover improve the novel class location module by making it more versatile to the developing stream, and empowering it to distinguish more than one novel class at the same time.
Key Words: Datastream, concept-evolution, novel class, outlier