The Impact of Partial Balance of Imbalanced Dataset on Classification Performance
Qing Li, Chang Zhao, Xintai He, Kun Chen, Runze Wang
Abstract
The imbalance of network data seriously affects the classification performance of algorithms. Most studies have only used a rough description of data imbalance with less exploration of the specific factors affecting classification performance, which has resulted in difficulty putting forward targeted solutions. In this paper, we find that the impact of medium categories on classification performance cannot be ignored, and therefore propose the concept of partial balance, consisting of Class Number of Partial Balance (β) and Balance Degree of Partial Samples (μ). Combined with Global Slope (α), a parameterized model is established to describe the difference of imbalanced datasets. Experiments are performed on the Moore Dataset and CICIDS 2017 Dataset. The experiment’s results on Random Forest, Decision Tree and Deep Neural Network show increasing α is a conducive step in the performance improvement of minority classes and overall classes. When β of dominant categories increases, that of inferior classes decreases, which results in a decrease in the average performance of minority classes. The lower μ is, the closer the sample size of medium classes is to the minority classes, and the better the average performance is. Based on the conclusions, we propose and verify some basic strategies by various classical algorithms.