Classification of Concept Drifting Data Streams Using Adaptive Novel-Class Detection
Aparna Yeshwantrao Ladekar, Mrityunjay Joshi · International Journal of Computer Engineering in Research Trends · 2016
In data stream classification there are many problems observed by the data mining community.Four major problems are addressed, such as, concept-drift, infinite length, feature-evolution and concept-evolution.Concept-drift occurs when underlying concept changes which is common in data streams.Practically it is not possible to store and use all data for training purpose whenever required due to infinite length of data streams.Feature evolution frequently occurs in many text streams.In text streams new features like words or phrases may occur when stream progresses.New classes evolving in the data stream which occurs concept-evolution as a result.Most existing classification techniques of data stream consider only the first two challenges, and ignore the latter two.Classification of concept-drifting data stream using adaptive novelclass detection approach is used to solve concept-drift and concept-evolution problem where novel-class detector is maintained with classifier.Novel-class detector is more adaptive to the dynamic and evolving data streams.It enables to detect more than one novel-class simultaneously.This approach solves feature-evolution problem by using feature set homogenization technique.Experiments done on Twitter data set and got reduced ERR rate and increased detection rate as a result.This approach is very effective as compared with existing data stream classification techniques.