Özet:

In recent years, huge increase in the number of people using Internet accompanied massive amounts of human and machine generated data recently called Big Data, where handling it efficiently is a challenging task. Along with that, valuable information that can be extracted from this data to perform data-driven decision making has attracted increased attention both from industry and academia. One of the important tasks in knowledge extraction is the classification task. However, in some of the real-world applications, dataset is either inherently skewed or collected dataset has imbalanced class distribution. Imbalance class distribution degrades the performance of several classification algorithms which generally expect balanced class distributions and assume that the cost of misclassifying an instance from both of the classes is equivalent. To tackle with this so called imbalanced learning problem, several sampling algorithms has been proposed in the literature. In this study, we compare sampling algorithms with respect to their running times and classification accuracies obtained from running classifiers trained with the sampled datasets. We find out that classification accuracies of the over-sampling methods are superior to the under-sampling methods. Sampling times are found to be similar whereas classification can be done more efficiently with under-sampling methods. Among the proposed sampling algorithms, the ADASYN method should be the preferred choice considering both execution times, increase in the data size and classification performance. Keywords: Imbalanced Learning, Sampling Methods, Data Mining, Big Data

Anahtar Kelimeler: