Methods of Balanced Random Sets and Data Normalisation for Improvement of Classification Quality

Abstract

In many cases direct application of the standard classification models leads to poor quality of results. In this paper we consider two examples. The subject of the first example are popular imbalanced data «Credit» from the platform Kaggle. Standard function nnet (neural networks) in the program environment R is used as a classificator. This function is ignoring an important minority class. As a solution to this problem, we are proposing to consider a large number of relatively small and balanced subsets, where elements were selected randomly from the training set. The subject of the second example are famous data MNIST and standard function svm (support vector machine) in the environment Python. The necessity of normalisation of the original features is demonstrated.