SVM Based Frameworks for the Classification of Evolving Streams An Application to Fraudulent Data Streams

Abstract

Many challenges arise from evolving data streams with class imbalance and concept newlinedrift. The class imbalance arises when one class significantly outnumbers the other class newlinedegrading the output of the minor class. Concept drift occurs when the underlying newlineconcept changes over time. As the model was built using outdated concepts it will be newlinedifficult for it to adapt to the current concept which reduce accuracy. It is difficult to newlineidentify concept drift in an imbalanced evolving stream. The performance of online newlinelearners is hampered by both concept drift and class imbalance either separately or in newlinecombination. In real-world applications, full access to class labels is impossible as it is newlineexpensive and time consuming to obtain ground truth. Detecting a drift in partially newlinelabelled data is difficult due to increase in false alarms. The goal of this work is to create adaptive classifiers that can handle concept drift and class imbalance in supervised and semi-supervised environments. newlineThe algorithms Online Oversample based Large Scale Support Vector Machine newline(OOLASVM) and Weighted Online Oversample based Large Scale Support Vector newlineMachine (WOOLASVM) are two variants proposed for supervised environments. They newlineare based on oversampling techniques. Extensive tests on different datasets show that the performance of the proposed methods is more for high class imbalances newlineFor identifying drifts in partially labelled data, a technique named Kolmogorov-Smirnov newlineArea Under Curve (KSAUC) is proposed. It uses a two-layer drift detector strategy with newlineunsupervised and supervised estimations. The proposed algorithm predicted the drifts newlinewith less false alarms newline