Online Approach to Handle Concept Drifting Data
Streams using Diversity
Parneeta Sidhu and Mohinder Bhatia
Division of CoE, Netaji Subhas Institute of Technology, University
of Delhi, India
Abstract: Concept
drift is the trend observed in almost all real time applications. Many online
and offline algorithms were developed in the past to analyze this drift and
train our algorithms. Different levels of diversity are required before and
after a drift to get the best generalization accuracy. In our paper, we present
a new online approach Extended Dynamic Weighted Majority with diversity (EDWM)
to handle various types of drifts from slow gradual to abrupt drifts. Our
approach is based on the Weighted Majority(WM) vote of the ensembles containing
different diversity levels. Experiments on the various artificial and
real datasets proved that our proposed ensemble approach learns drifting
concepts better than the existing online approaches in a resource constrained
environment.
Keywords: Online learning, ensemble,
concept drift, data streams, diversity.
Received October 29, 2013;
accepted December 16, 2014