Reduction Strategies to Tackle Class Imbalance in Datasets

Abstract

Banking, retail, financial, scientific and telecommunications and various other sectors have newlineall been using data mining technologies, for processing massive amounts of data measured newlinein zeta bytes. While this massive amount of data is useful, datasets have to be processed newlineeffectively to perform predictive and inferential forecasts for a target population. The Class newlineimbalance, where there are fewer instances of a class than the number of instances in other newlineclass/classes in a dataset has posed challenges to the traditional classifiers. Traditional newlineclassifiers fail to handle the imbalanced datasets due to inherent assumptions made in newlinedesigning them. The distribution of classes within the dataset has a direct impact on the newlineclassifier/model performance. One of the proven practices to address this problem is to newlinebalance the classes in the training data sets. Main goals of the balancing are increasing newlinesensitivity, selecting representative samples from the majority class, maintaining newlinetrade-off between Majority Class and Minority Class prediction rates. newlineThis thesis aspires to address the inadequacies of data science models caused by class newlineimbalance problem using data reduction strategies. In order to achieve these goals, five newlinetechniques Ensemble based Classification using Small Training sets (ECST), Centroid newlineBased Grouping (CBG), Quartile based Under Sampling (QUS), Mahalnobis distance newlinebased Centroid based Undersamplig with Filter (MahalCUSFilter) and Simulated Annealing newlinebased Under Sampling (SAUS) are proposed here. ECST focuses on getting good newlinesensitivity by generating small balanced training sets and using ensemble classification newlineto produce outcomes specified in the goals. CBG generates prototypes(artificial samples) newlinefrom the original training set and uses the Lp distance metric to classify test set samples newlinein order to account for neighbourhood space. The QUS algorithm groups each negative newlineinstance with one of the five quartiles. This is how negative samples from the full negative newlinetraining distribution are selected to bu