Mining of Data Stream In Presence of Concept Drifts
Loading...
Date
item.page.authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In stationary (or static) data streams, the underlying data distribution of the data stream
newlineremains uniform throughout the stream. Conversely, in non-stationary data stream the
newlineunderlying data distribution of the data stream changes very frequently with respect to
newlinetime.
newlineHowever, since, in the real world, most of the applications are inherently dynamic in
newlinenature and produce data streams with non-uniform underlying data distribution, hence,
newlinethis non-uniformity increases challenge in data stream mining with classical batch
newlinelearning approaches of data mining and machine learning directly. The data streams
newlinewhose underlying data distribution changes over time are commonly referred as concept
newlinedrifting data streams and the change in the underlying data distribution is called concept
newlinedrift.
newlineInspired from the above discussions and previous studies, this dissertation investigates
newlinecertain major issues in the existing state-of-art approaches of concept drifting data
newlinestream mining by performing empirical evaluations on some standard artificial and real
newlineworld bench mark datasets in order to explore preferable combinations of base
newlineclassifiers and concept drift dealing algorithms for all types of concept drifts. Instead
newlineof utilizing only classical learning and testing strategies, this work includes some new
newlinelearning strategies and performance measure matrices for facilitating the online
newlinelearning and run time accuracy monitoring.
newlineMoreover, in this thesis, a novel approach is proposed called Age Decay Accuracy
newlineWeighted (ADAW) Ensemble Architecture for Learning in Drifting Data Streams.
newlineADAW is empirically evaluated on some artificial drifting data stream generators and
newlinecompared its performance with nine well-known state-of-the-art existing methods. The
newlineexperimental results show that our proposed ensemble model ADAW outperforms over
newlinethe existing ensemble methods and also performs equally well for all kinds of concept drifts.
newline