Improving instance selection methods for big data classification

The explosion of data in many application domains leads to a new term called big data. While the big data volume rapidly exceeds, the capacity and processing capabilities of contributed data mining algorithms are not effective. The instance selection methods become a mandatory step prior to applying data mining algorithms. Instance selection methods scale training set down by eliminating redundant, erroneous, and unrelated instances. Recently, instance selection methods have improved to work on big data sets by splitting training data into disjoint subsets and applying instance selection methods on individual subsets. However, these improved methods have a variable performance in the degree of reduction rate and classification accuracy. In this work, we propose an operational and unified framework to balance between reduction rate and classification accuracy. It starts with splitting a training set into class-balanced subsets to analyze the impact of the splitting process on the performance regarding the reduction rate and classification accuracy. It then applies two different instance selection methods on each subset. We implement and test experimentally the framework using two standard data sets. With the random splitting process as a benchmark, the results prove that the class-balanced splitting process is preferred regarding the classification accuracy criterion. The results also depict that the combination of two instance selection methods remarkably reduces the performance variability.