A matched-pair comparative study on classification of data streams with concept drift

Increasingly, the Internet of Things (IoT) realms, social media applications, hardware devices and etc. generate data at an astonishing rate. These continuously incoming data from heterogeneous sources is referred to as data stream. Data stream mining is not only an urgent trend topic but also entangled. The dynamic, unbounded, high-dimensional and rapidly evolving structure of stream data rendered traditional data mining techniques insufficient. The real-world applications particularly involve real-time streams in an evolving environment; network intrusion detection, weather prediction, spam e-mail filtering, fraud detection, etc. Online learning derives information from the large volume of stream data, usually affected by the changes in the underlying distribution; often in unforeseen ways. This phenomenon, called as concept drift, makes formerly learned models insecure and imprecise in classification manner. As a handling method, concept drift detectors attempt to estimate the position of concept drift in data streams in order to substitute the base learner after drift has occurred and try to improve overall accuracy. This study propose a matched-pair comparison between 13 drift detectors [DDM, EDDM, ADWIN, CUSUM, GMA, PageHinkley, ECDD, HDDMA, HDDMW, SEQDRIFT2, STEPD, RDDM, and SEED] and 8 classifiers as base learners [Naive Bayes(NB), HoeffdingTree(HT), HoeffdingOptionTree, Perceptron(PR), OzaBagASHT, OzaBagADWIN, Decision Stump(DS), and k-Nearest Neighbor(kNN)]. In parallel with the aim of thesis, pairs (classifier, detector) with higher accuracy scores are recommended. The experimental evaluation is conducted on Massive Online Analysis (MOA) software.