Search

The hierarchical dimensionless method can effectively remove the effect of different index dimensions, and solve imbalanced data distribution and low discrimination caused by anomalous index values. However, when using this method, it is necessary to artificially specify the number of partition intervals so that the dimensionless results are interfered by human factors and lose objectivity. To solve this problem, a dimensionless method of density hierarchy is proposed considering the distribution characteristics of raw data. This method divides the interval according to the density of data distribution, objectively determines the hierarchical series, and takes into account the advantages of the hierarchical dimensionless method. The calculation is comparatively simple and reduces human factors. In addition, through the stochastic simulation method, it is found that the method has good anti-interference to outliers, and the balance of dimensionless results is affected by the scale of raw data. [ABSTRACT FROM AUTHOR]

Aiming at the problems of the identification precision of traditional classifier was low for the minority class sample and the traditional under-sampling method was easy to lose information, the paper proposed an under-sampling method US-DD based on data density distribution, which could divide data into high density data cluster and low density data cluster. The two kinds of data were not only different on the sample quantity, but also were different on the influence of data classification. Therefore, it could divide the data set using the data density, performed different re-sampling strategy for the data cluster with different density, and achieved the purpose of improving data balance. Through selecting six IJC1 data sets, the experiment results show that the US-DD method is effective for unbalanced data classification, and can effectively improve the recognition performance of the classifier for the minority class by comparing with the method of the random under-sampling and KNN-NearMiss. [ABSTRACT FROM AUTHOR]